Entity Confidence Registered Trade Mark

Author: Tony Lord

  • Is your website locking AI out?

    Most conversations about AI visibility assume the problem is that you don’t have enough of the right content. But sometimes the reason is that your website is quietly turning away the very software that needs to read it. Your business isn’t invisible for want of something to say. But perhaps it has been set up in a way that has locked the AI visibility entrance.

    This is a better problem to have, because unlike most of AI visibility, it is fixable in an afternoon.

    What we found

    In a dry run of our AI Visibility Check across 45 businesses in one local-service sector, 7 of the 45 websites (about one in six) could not be read by the check at all. [1] Instead of any result, the owner would just see a message saying we couldn’t complete the assessment.

    Two of those seven were down to out-of-date web addresses in our own records — a real business with a working site we simply had the wrong link for. Set those aside, and roughly one in nine legitimate businesses had a website our check genuinely could not read. [2] The reasons:

    • Most were the site actively refusing the automated request — a “403 Forbidden,” which is a website’s way of saying “no robots allowed.” Three of the five. [3]
    • One had a security-certificate problem that stopped the page loading cleanly. [3]
    • One was built entirely in JavaScript with no plain text for an automated reader to pick up. [3]

    Different causes, same outcome: software arrives at the page and comes away with nothing.

    Why a locked door makes you invisible

    When an AI assistant answers a question like “best plumber in [locality]” or “who should I use for X near me,” it relies on what it can actually read from the live web. If your site returns “no robots allowed” to an automated visitor, the assistant can’t read your own words about who you are, what you do, or where you are — so it has nothing of yours to work with, and reaches for competitors it can read instead. You can have the best website in your market and still be absent from the answer, because the answer is written from competitor pages the machine was allowed to open. [4]

    It is worth being precise here: our check is not the same software as any given AI engine’s, and a site that blocks one automated visitor doesn’t necessarily block every one. But a blanket “no robots” rule is blunt — it tends to turn away the legitimate readers along with the nuisance ones. If your site is blocking automated traffic wholesale, the crawlers behind AI answers are very likely on the wrong side of that wall too. [4]

    How to find out, and what to do

    You don’t need to fix any of this yourself. There are three questions to put to whoever looks after your website — an agency, a developer, or the person who set it up:

    • “Does our site block bots or automated traffic?” Many sites switch this on — through a security plugin, a firewall setting, or an aggressive bot-blocker — without anyone deciding the AI assistants should be shut out too. It is often a setting, not a rebuild.
    • “Is our security certificate valid and not mismatched?” A certificate warning doesn’t just unnerve human visitors; it can stop an automated reader loading the page at all.
    • “Is our content readable without running scripts?” A site built entirely in JavaScript with little plain text can look perfect to a person yet be blank to a reader that doesn’t execute code.

    Then re-check. Once the door is open, the same content that was invisible can start being read — and used.

    The reason this matters more than it first appears: a business with thin content has a long job ahead of it. A business whose website is simply shutting AI out has a short one. It is not that AI has judged you and found you wanting. It is that AI never got to see you at all — and that is among the most fixable problems in this whole field.

    —

    These figures come from EC’s own AI Visibility research: a single grounded AI engine, a small sample, and a live snapshot of the web that changes over time. They illustrate a pattern, not a fixed market rate. Method: entityconfidence.ai/research-methodology/ · Corrections and right of reply: entityconfidence.ai/corrections-and-right-of-reply/

  • What the CMA’s private dentistry study means for your practice

    By Hedley Basford, 30th September 2026. This article reflects the position of the study at that date.

    On 5th March 2026 the Competition and Markets Authority (CMA) launched a market study into private dentistry in the UK, a sector it describes as worth £8 billion. The study followed a request from the Chancellor, and it will shape how patients find, compare, and choose private dentists for years to come. Here is what the CMA is looking at, what it has heard so far, and what it means for a private practice.

    Why the CMA is looking at private dentistry

    Private care now accounts for the majority of the UK dental market. The CMA cites figures showing that private services made up 69% of the market in 2025, and that one in five people in Great Britain used private dental care in 2024. It adds that research suggests the shift is partly driven by difficulty accessing NHS treatment.

    The CMA has been clear that the study is not a criticism of clinicians or the care they provide. It is an examination of how the market is working for patients, including whether they have the information they need to make good choices.

    What the study covers

    The CMA has confirmed five themes: consumer journey and choice, competition, conduct, regulatory frameworks, and market outcomes. The first of these is the most relevant to how practices are found: how patients access, assess, and act on information when choosing a dentist, and how easily they can switch between practices.

    Under competition, the CMA is looking at how practices compete to win and keep patients. Some market participants and representative bodies told it that the factors influencing competition include the availability of appointments, opening hours, the range of treatments offered, ease of online booking, and the quality of premises.

    What the CMA has heard so far

    Nearly 1,000 dental professionals and more than 1,800 consumers responded to the CMA’s call for views. The CMA notes that these responses may not be representative, and it is running separate, statistically robust research alongside them. Even so, several themes stand out for practice owners:

    • Some patients reported difficulty comparing treatment options and their cost, and information in practices and on practice websites was thought to be important.
    • Independent practices make up around 65% of UK dental practices, and some respondents pointed to the difficulty independents face in competing with larger groups, including because of lower marketing budgets.
    • One dental professional told the CMA that private dentistry is competitive, “especially now as more and more practices are leaving the NHS”.

    What happens next

    The CMA expects to publish its emerging thinking between October and November 2026 and will invite views on it. Its final report is due by 4th March 2027. The CMA has said that possible outcomes include recommendations to improve transparency and help patients make good choices.

    What it means for your practice: our view

    Patients are being encouraged to compare. The first of the CMA’s tips for private dental patients is to compare what is available: to search online and ask friends and family where they go, and to consider whether a practice offers NHS and private treatment, what it costs, and whether its facilities and opening hours suit them.

    In our view, AI-powered search engines are becoming one of the places that comparison happens. When a patient asks one to recommend a private dentist, it names a small number of practices, drawing on the information it can find about each. Three practical steps follow:

    • Make your practice easy to compare. Keep your services, your prices, and your opening hours clear and current on your website. The General Dental Council’s standards already require a simple price list in your reception or waiting area, and clear information on prices on your website.
    • Keep your information consistent everywhere. Your practice name, address, and services should match across your website, your Google Business Profile, and your CQC registration.
    • Watch the CMA’s emerging thinking this autumn. Any recommendations on transparency are likely to affect what patients expect to find before they choose a practice.

    Entity Confidence shows private dental practices which practices AI recommends in their town, where their own practice sits, and what will strengthen its position, month by month. We work with one practice per town. Check whether your town is available.

    Sources

    Contains public sector information licensed under the Open Government Licence v3.0.

  • Independent Coverage Is What Gets You Recommended by AI

    If you spend money on PR, there is an awkward possibility worth checking: the coverage you are working hardest to get may not be the coverage that decides whether an AI assistant recommends you.

    Muck Rack compared the journalists PR teams pitch most often against the journalists AI engines actually draw on for a given brand. The overlap averaged 2%.

    That is not a small inefficiency. It suggests most PR effort is aimed somewhere other than where AI systems are looking — which matters more every year, as more of your customers start their search by asking an assistant rather than typing into Google.

    Coverage You Didn’t Pay For Is Doing Almost All the Work

    Start with the headline finding, because it is stark.

    Muck Rack’s continuing study of what AI systems cite — more than 25 million links across ChatGPT, Claude and Gemini, covering seventeen industries — found that earned media accounts for 84% of all AI citations, while paid and advertorial content accounts for 0.3%. “Earned media” means coverage you did not pay for and cannot control: a journalist’s article, a trade publication’s round-up, an analyst’s assessment. The figure has held between 82% and 89% across three editions of the study since July 2025.

    Nought point three per cent for everything you can simply buy.

    Academic work points the same way. Researchers at the University of Toronto took 300 consumer-electronics queries and classified where each system’s sources came from. Claude drew 65% of its sources from independent media and review outlets, GPT-4o 57%, Perplexity 50%, Gemini 46% — against Google’s 41%. For the queries that matter commercially — best-of lists, comparisons, “which should I choose” — the AI engines converge on earned sources at between 59% and 86%.

    So when someone asks an AI which supplier to use, the answer is assembled largely from what other people have written about you. Not from your website.

    Why Editors Count for More Than Your Own Pages

    The mechanism is not mysterious. An AI system has no way to verify what you say about yourself. It can check whether anyone else says the same thing.

    When a trade publication names you in a piece, the decision to include you was made by someone with no stake in your marketing. That is evidence of a kind your own website cannot produce, however well written it is.

    Ahrefs measured this across 75,000 brands, looking at which signals track with a brand’s visibility in Google’s AI Overviews. Mentions of a brand across the web scored 0.664. Backlinks — the currency of search optimisation for two decades — scored 0.218. Ahrefs is careful to say this shows association rather than proven cause, and so are we. But the direction is consistent with everything else here: being talked about tracks visibility more closely than being linked to.

    Which Outlets, Not How Much Coverage

    This is where that 2% figure becomes useful rather than merely alarming.

    It does not mean PR does not work. It means most PR programmes are aimed at publications chosen for reach, prestige or relationships — sensible criteria that happen not to be the ones AI systems respond to.

    AI citation is concentrated. In the categories the Toronto researchers examined, the sources AI systems leaned on were overwhelmingly specialist review and comparison publications rather than the national titles a PR brief instinctively reaches for.

    Knowing which publications carry weight in your category turns a vague “more coverage” brief into a targeted list. That is a more valuable piece of information than an extra month of outreach.

    It Has to Keep Going

    Coverage decays.

    Muck Rack found half of all AI citations come from articles published within the last eleven months, and about 4% from the previous week alone.  The Toronto researchers measured median source ages by engine: Claude 62 days, GPT-4o 80, Perplexity 90, against Google’s 130 — AI engines drawing on material roughly 31–70% newer than Google does.

    That does not make older coverage worthless; the remaining half of citations spans roughly three further years, and long-lived reference pages do a lot of quiet work. But a campaign you ran eighteen months ago is not carrying you now. This is a drumbeat, not a launch.

    Being Cited Is Not the Same as Being Named

    There is a complication worth understanding before you judge whether any of this is working.

    A large share of AI citations are what the industry calls “ghost citations” — your page is linked as a source, but your brand name never appears in the answer the customer reads. A study of 3,981 brand appearances across four AI engines found that 74.9% drew a citation but only 38.3% got the brand named in the answer text. In 61.7% of appearances the link was there and the name was not — which is 82% of all the citations earned.

    The practical consequence is uncomfortable. Your content can be doing the work of informing the answer while a competitor gets the recommendation.

    The fix is partly structural. Coverage you secure should carry your name in the sentence a machine would extract, not buried in surrounding context — which in practice means comparative and evaluative pieces, where naming is intrinsic, rather than purely informational ones.

    What This Means for You

    If AI assistants are not recommending your business, the most likely fix is not on your website.

    Three things follow. First, the question worth asking your PR agency is not “how much coverage did we get?” but “which publications in our category do AI systems actually draw on, and are we in them?” Those are different questions with different answers. Second, budget for continuity rather than campaigns — a steady presence beats an annual spike, because the systems weight recent material heavily. Third, check whether the coverage you already have names you in a way a machine can extract, or merely links to you.

    None of this requires a large budget. It requires aiming at the right targets, which is mostly a matter of knowing what they are.

    Where you stand can be measured rather than guessed at: whether AI tools name your business when someone asks for what you do, how that compares with your competitors, and what in your existing online reputation is producing that answer.

    That is what we do at Entity Confidence.

  • Why ChatGPT Still Needs Google

    One of the more persistent myths in AI search is that ChatGPT and Google are in competition — two separate systems fighting for the same users, pulling from separate pools of content. The practical version of this belief shows up in marketing strategy discussions: if AI search is taking over, does SEO still matter? Should budgets shift away from search optimisation and toward something else?

    The answer is no, and the reason is structural. ChatGPT is not independent of Google’s infrastructure. It is, to a significant degree, built on top of it.

    The 88% Finding

    Ahrefs analysed 1.4 million ChatGPT prompts and traced where the cited pages actually came from. The search channel — content retrieved via the web search index — accounts for 88% of all ChatGPT citations. News accounts for most of the rest. Reddit, YouTube, and academic sources together contribute less than 3%.

    This is not a peripheral finding. It is the foundational architecture of how ChatGPT browses. When a user asks ChatGPT a question that triggers web retrieval, the system queries a search index, assembles candidate URLs, filters them for relevance, reads the most promising pages, and cites selected content. The index it queries is built substantially on the same crawl infrastructure that powers web search.

    A brand not present in that index is not retrievable. It does not matter how well the brand’s content is structured, how authoritative its coverage, or how precisely its messaging matches a query. If the page is not indexed, it is invisible to the retrieval pipeline.

    The operational implication: making a page indexable and crawlable by search engines is not SEO housekeeping. It is the minimum requirement for AI citation eligibility.

    Indexed, Not Necessarily Ranked

    The 88% figure is sometimes misread. It does not mean that ChatGPT cites pages because they rank well in Google. The relationship is looser than that.

    Research across multiple sources finds that 80% of LLM citations do not rank in Google’s top 100 for the specific query being answered. These pages are indexed — they are in the database — but they are not necessarily winning on traditional search ranking signals for the particular question being asked.

    The distinction matters. What ChatGPT needs from Google’s infrastructure is discovery and access: a mechanism to find pages and retrieve their content. It does not need those pages to have won Google’s ranking competition. A page that sits at position 47 for a given query, or that ranks well for related queries but not this exact one, can still be retrieved and cited by ChatGPT if it passes the semantic relevance filters.

    The threshold is being indexed, not being ranked. But the two are related in practice. Pages that are not indexed are invisible. Pages with severe technical issues — slow load times, blocked crawl paths, thin content signals that discourage indexing — are poorly represented even when nominally indexed. And pages that rank well for a topic tend to be well-indexed, freshly crawled, and more likely to appear in AI retrieval sets.

    The correct frame is not “rank #1 or be invisible to AI” — it is “be a legitimate, accessible, well-structured presence in the web index, and AI can reach you.”

    Google AI Overviews: A Tightening Relationship That Has Since Loosened

    The dependency between AI citation and search performance was, until recently, the tightest anywhere in AI search — and it has since weakened considerably.

    In July 2025, 76% of AI Overview citations came from pages already ranking in Google’s top 10. For Google’s own AI-generated answers, traditional search ranking was not a rough correlation — it was the dominant selection mechanism. The E-E-A-T signals that determine page ranking were directly inherited by AI Overview citation selection. Branded web mentions correlated with AI Overview citation at 0.664 — the strongest single signal measured in Ahrefs’ original research.

    That figure has since collapsed. Ahrefs’ Brand Radar analysis of 863,000 SERPs and 4 million AI Overview URLs found that by January 2026, the top-10 share of AI Overview citations had fallen to 38% — driven by query fan-out, the mechanism by which AI Overviews decompose a single question into multiple sub-queries and cite from a far wider, more dispersed pool of pages. Our Confident and Wrong series covers this collapse and its implications in detail.

    This does not mean Google AI Overviews and Google Search have become architecturally unrelated — AI Overviews are still an interface layer built on top of Search’s infrastructure, and a page still has to be indexed and competitive to enter the retrieval pool at all. But the direct inheritance from ranking to citation that held in mid-2025 no longer holds to the same degree.

    For brands targeting visibility in Google’s AI products specifically, ranking well in Google Search remains necessary but is no longer close to sufficient — a top-10 ranking now predicts AI Overview citation at roughly the rate of a coin flip, not the near-certainty it represented six months earlier.

    Where the Dependency Breaks Down

    The search-channel dependency is not uniform. Two scenarios produce exceptions to the 88% pattern.

    The training-mode exception. When a user asks ChatGPT a question it can answer from pre-trained knowledge — without triggering any retrieval — the search channel plays no role. The model draws from parametric knowledge built during training, shaped by the web’s cumulative discussion of a topic. For well-known brands with strong training-data presence, a significant share of AI mentions may be training-mode responses that retrieve nothing and cite nothing.

    This is not an opportunity to avoid search investment. It is a separate channel (documented in detail elsewhere), and it is largely inaccessible to short-cycle optimisation — it is determined by years of web presence, not recent content. For the niche and mid-market brands that make up most commercial AI search competition, the pre-training channel is weak by definition. The AI has limited knowledge of them at training time, so it relies on retrieval — and retrieval runs through the search index.

    The Perplexity exception. Perplexity’s citation pool overlaps with ChatGPT’s by less than 1%. Its architecture weights Reddit more heavily (46.7% of top Perplexity citations) and draws from a different mix of authoritative sources. It is also, contrary to a widely repeated claim, the least recency-biased of the major engines: of the pages Perplexity drew on in a study of roughly 47,000 citations tracked between March and June 2026, 65% had been updated within the previous year, against Gemini’s 78% and ChatGPT’s 73%. Optimising for ChatGPT citation does not automatically produce Perplexity visibility.

    This matters for brands choosing where to invest. ChatGPT and Perplexity are not interchangeable targets, and a strategy built entirely around the search-channel dependency will be less effective for Perplexity than for ChatGPT or Google AI Overviews.

    What This Means for SEO Investment

    The 88% finding does not mean SEO and AI search are identical — it means SEO is the infrastructure layer on which most AI citation depends. Three practical conclusions follow from this.

    Search indexability is non-negotiable. Any brand whose pages have crawl issues, thin-content penalties, or poor technical foundations is not just underperforming in search — it is actively reducing its AI citation eligibility. Technical SEO is not a legacy practice in an AI-first world. It is the precondition for entering AI retrieval pipelines.

    Search ranking improves AI citation probability, but the relationship is not linear. A page that moves from position 12 to position 3 in Google Search will be more likely to appear in AI retrieval sets — not because AI citation tracks ranking directly, but because high-ranking pages are fresher, better crawled, and more likely to pass AI semantic relevance filters. The investment that improves ranking also tends to improve citation eligibility, through overlapping mechanisms.

    AI citation requires an additional layer that search ranking alone does not provide. Getting indexed and ranked makes a page reachable. What determines whether it is actually cited is a second filter: semantic relevance to the AI’s internal sub-questions, content structure (direct answers, clear headings, attributed data), and source credibility (authoritative third-party coverage far outperforms brand-owned content). These are not ranking factors in the traditional sense — they are citation selection factors that apply after the retrieval pipeline has already found the page.

    The sequence is: be indexed → be retrievable → be selected for citation. Search investment covers the first two stages. AI-specific content and authority work covers the third.

    The Strategic Error to Avoid

    The error is treating AI search and traditional search as substitutes — allocating budget to “AI SEO” while reducing investment in the infrastructure that makes AI citation possible in the first place.

    ChatGPT’s 88% search-channel dependency is not a transitional feature that will disappear as AI search matures. It reflects a structural reality: AI retrieval systems need curated, authority-evaluated, freshness-maintained content databases to draw from, and the web’s search index is the most comprehensive one that exists. Google built that index over two decades. AI systems are using it because there is no better alternative.

    The brands that will be cited reliably in AI answers are the ones that are indexed, crawlable, authoritative, and structured for extraction — then additionally covered by the earned media and community presence that AI systems use to evaluate authority. The first set of requirements is not new. They are what good SEO has always required.

    The second set of requirements is the genuine addition. But it sits on top of the first set, not beside it.

    Sources: Ahrefs 1.4M ChatGPT prompt study (Why ChatGPT Cites One Page Over Another); Ahrefs AI Overview brand correlation research (July 2025 baseline); Ahrefs Brand Radar, AI Overview citation analysis (February 2026, data collected January 2026); Seer Interactive, “Content Recency’s Impact on AI Visibility in 2026” (July 2026); Chen et al. arXiv:2601.16858 (University of Toronto), EDBT/ICDT 2026 Workshops; BrandFeatured AI ranking factors cross-platform divergence data.

    Revised and republished 24 September 2026

  • Updating Your Best Pages Beats Publishing New Ones

    Date: 2026-09-24

    If the pages on your website haven’t been touched in a couple of years, AI tools are quietly less likely to use them when someone asks for a recommendation in your line of work.

    That sounds like an argument for publishing more. It isn’t. The research points somewhere more useful and a good deal cheaper: the pages you already have, kept up to date, do more for you than a stream of new ones. Most businesses have this backwards, which is why it’s worth setting out what the evidence actually shows.

    “Publish fresh content” has been the standing advice for two years now. What nobody attached to it was a number. Fresh compared to what — last month, last year, three years ago? Without that, it’s a vague instruction rather than something you can act on. There are now numbers, from three large independent studies, and they tell a more interesting story than the advice does.

    AI tools draw on noticeably newer material than Google does

    Researchers at the University of Toronto measured how old the pages were that each system actually drew on when answering buying-type questions. They ran the same set of questions through Google and through three AI assistants, then checked the publication date of every page each one used.

    The typical age of the material they leaned on:

    Where the answer came fromConsumer electronicsCars
    Claude62 days148 days
    ChatGPT80 days162 days
    Perplexity90 days217 days
    Google130 days493 days

    Read across the bottom row first. In a slower-moving market like cars, the typical page Google surfaced was nearly sixteen months old. The AI assistants were working from material five months old. Across both markets, AI answers drew on content between roughly a third and 70% newer than Google’s.

    The gap matters because of what it implies about the competition. Older pages aren’t banned — but they’re up against a pool of material that skews much younger than the one you’re used to competing in on Google. And the slower your sector moves, the wider that gap gets. If your competitors are sitting on pages they last touched in 2023, that’s an opening rather than a threat.

    Most content AI draws on is still years old

    Here’s the finding that stops this becoming a panic about publishing schedules.

    Ahrefs looked at nearly 17 million links cited across seven different AI and search platforms. On average, the pages AI tools drew on were 2.9 years old. Google’s ordinary results averaged 3.9 years. So yes — AI leans newer, by about a year. But the typical page it uses is still the best part of three years past its publication date.

    Both findings are true, and the difference between them is worth understanding rather than averaging away. The Toronto study measured the midpoint age in two fast-moving consumer markets where publishers refresh constantly. The Ahrefs figure is an average across every topic, and a handful of very old, very authoritative pages pull an average up in a way they don’t pull a midpoint up.

    What survives both readings is the same: recency helps, and it helps more than it used to — but a three-year-old page that genuinely answers the question is still being used every day. Age alone doesn’t disqualify you.

    Updating an old page counts as fresh

    This is the finding that changes what you should actually do on Monday morning.

    Seer Interactive tracked 7,683 pages that AI tools drew on across roughly 47,000 mentions between March and June 2026. Three in four of those pages had been updated within the previous year. Judged by when they were last updated, nearly three-quarters looked current.

    Then they looked at the same pages by original publication date. On that measure, fewer than half were recent. More than a quarter of the pages that qualified as “fresh” had first been published more than two years earlier.

    In other words: most of the fresh content AI tools are using is not new content. It’s old content that somebody maintained.

    That is a much easier instruction to follow than “publish more”. You don’t need a content team. You need to know which of your existing pages matter — your main service pages, your pricing or process explanations, the questions customers actually ask you — and to keep them genuinely current.

    The word “genuinely” is doing real work there. Changing a date stamp without changing anything else is not an update, and the author of the Ahrefs study warns explicitly against that. What counts is substance: a price that’s right, a service you no longer offer removed, a new question answered, a case or example from this year rather than three years ago.

    How fresh is fresh enough?

    There’s no single threshold, and anyone who gives you one is guessing. But the evidence supports a reasonable rule of thumb.

    Within the last year is the broad working window. Of the pages AI tools drew on in the Seer study, Google’s Gemini took 78% from the previous twelve months, ChatGPT 73%, and Perplexity 65%. Separately, Muck Rack’s analysis of more than 25 million links across ChatGPT, Claude and Gemini found that more than half of the news coverage cited was published within the past year, with the volume dropping off sharply after the first six months.

    So: for the handful of pages that genuinely matter to your business, once a year is the minimum, and twice is better. For a fast-moving sector, or for a page carrying prices, availability or anything else that dates quickly, work to a shorter cycle. That’s it. It doesn’t need a dashboard.

    What this means for you

    If you have been feeling guilty about not publishing enough, you can stop. The more useful question is whether the pages you already have are still accurate, and when you last checked.

    Pick the five or six pages that would matter most if a prospective customer read them tomorrow. Look at when each was last meaningfully changed — not restyled, changed. If any of them describes a price you’ve since altered, a service you’ve stopped offering, or a world that has moved on, that is a page currently working against you in two places at once: with the customers who read it, and with the AI tools deciding whether to put you in front of the next one.

    There’s a wider point behind this. Whether AI tools recommend your business isn’t a mystery and it isn’t luck. It can be measured — how often you’re named when someone asks for a business like yours, how that compares with your competitors, and what in your wider online reputation is driving the answer either way. Content age is one input among several, and on its own it will never be the whole story.

    Finding out where you currently stand is what we do.

    Sources: Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas (University of Toronto), “Navigating the Shift: A Comparative Analysis of Web Search and Generative AI Response Generation”, EDBT/ICDT 2026 Workshops; Ryan Law, “Do AI Assistants Prefer to Cite Fresh Content?”, Ahrefs, July 2025; Seer Interactive, “Content Recency’s Impact on AI Visibility in 2026”, July 2026; Muck Rack Generative Pulse, “What Is AI Reading?”, May 2026.

  • AI Traffic Is Better Traffic

    The people who reach your website from an AI tool are turning out to be some of the best customers you can get.

    That runs against the usual story, so it’s worth explaining. When someone asks ChatGPT, Google’s AI answers, Perplexity or a similar tool a question, they often get everything they need on the spot and never visit the website the answer came from. This is called “zero-click”, and it is usually treated as bad news: if people get their answer without visiting you, fewer of them land on your site, and fewer of them buy.

    That worry is real. But it is only half the picture. The visitors who do click through from an AI answer behave very differently from ordinary web traffic — and the difference has been moving firmly in your favour.

    Visitors who arrive from AI now buy more, not less

    The clearest evidence comes from Adobe, which tracks well over a trillion visits to US retail websites, so this is a very large sample rather than one company’s anecdote.

    A year ago, the picture looked exactly as pessimists predicted. In March 2025, people arriving from AI tools were 38% less likely to buy than visitors from every other source. AI seemed to be sending browsers, not buyers.

    Then it reversed. By March 2026 the same measure had flipped the other way: AI visitors were 42% more likely to buy than everyone else — a record at the time. And it kept climbing. By May 2026, AI-referred shoppers were converting 54% better than non-AI traffic, while the volume of that traffic more than doubled year on year. On Prime Day 2026, visitors arriving from AI tools converted better than those from any other channel.

    Same measurement, opposite result, in a little over a year.

    Why the quality turned around

    Two things changed at the same time.

    First, people started trusting AI answers. Roughly two in three shoppers now say they believe AI tools give accurate results — up sharply from a year earlier. As that trust grew, the kind of person using AI to shop changed too: from curious early adopters trying something new, to mainstream buyers happy to act on what an AI recommends.

    Second, the questions people ask an AI are a different shape from what they type into a search box. An AI search query runs to about 23 words on average, against four or five for a traditional search. That matters more than it sounds. A four-word search is often idle browsing. A 23-word question is a decision in progress — someone who has already thought hard about their problem, knows roughly what they want, and is describing it in detail. By the time the AI points that person to you, they have effectively been pre-qualified by a conversation you never had to have. They arrive further along the path to buying than a visitor who found you by typing two keywords.

    The way those visitors behave once they land backs this up. Compared with everyone else, people arriving from AI tools spend about half as long again on the site, browse close to a quarter more pages, and engage noticeably more. These are people doing serious homework, not passers-by.

    The same pattern shows up everywhere else

    Adobe’s numbers are the largest, but not the only ones pointing this way.

    An analysis by Onely, drawing on Ahrefs data, found AI-referred visitors converting many times better than traditional search visitors. The same work found people arriving via Perplexity spending more than nine minutes on the sites they landed on, and returning to them 85% of the time. One software company reported that although AI sent it only half a percent of its total visits, those visits produced roughly one in eight of all its sign-ups.

    The pattern is consistent across every source: AI recommendation sends you fewer visitors, but far more of them are ready to act.

    This makes sense once you picture what has happened before the click. The person has described their situation to an AI assistant, been given a recommendation, and chosen to click through and look closer. The AI did the early research-and-shortlisting work that your website would otherwise have had to do for a cold visitor. What lands on your page is what’s left after that filtering: the genuinely interested.

    The honest trade-off

    None of this makes the zero-click worry disappear. Volume really does fall. Most people who get a useful AI answer about your business will never visit your site at all, and click-through rates from ordinary search have dropped as AI answers have spread. Both things are true at once: the number of visits falls, and the quality of the ones you do get rises.

    Whether that is good or bad for you depends on how you make money. If a handful of high-value customers drives most of your revenue — professional services, considered purchases, business-to-business, subscription software — then better-quality visitors can easily be worth more than the lost volume. If your model depends on high volumes of small, low-margin sales, higher quality may not make up for fewer visits.

    But one implication holds for every business, whatever the model: the number of visits in your website analytics is becoming an unreliable measure of how much AI search is affecting you. A business being recommended constantly in AI answers — shaping decisions, sending high-intent buyers, influencing people who never click at all — might see barely any change in its visit count while its commercial reality shifts underneath it.

    What to measure instead

    Your conversion figures are encouraging, but they only tell you what happened after the AI had already made its recommendation. By the time an AI-referred sale shows up in your analytics, the important moment has passed. You have no sight of the questions where you weren’t recommended, the ones where a competitor was, or the many conversations that shaped a buyer’s thinking before they visited anyone’s site.

    The measure that comes first is your AI Visibility — how often you actually get named and recommended when AI tools answer the questions your buyers are asking. That is the leading signal. It is fed by your Reputation: what the wider web says about you, because that is the evidence an AI draws on when it decides who to name. Conversion quality then confirms, after the fact, that the traffic was worth having when it arrived.

    Both matter, and neither replaces the other. But if the only thing you are watching is what lands in your analytics, you are seeing a small fraction of where AI search is already deciding whether people choose you.

    Sources: Adobe Analytics AI traffic reports (over 1 trillion US retail visits; Q1 2026 through May 2026 data, published June 2026; Prime Day 2026 update); Onely / Bartosz Góralewicz, “Zero-Search Discovery” (December 2025); Ahrefs (via Onely).

  • Why the area you ask about changes whether AI can find you

    One of the most common reasons a business looks invisible to AI is not the business at all. It is the question. Ask an assistant about your county and you can look completely absent. Ask it about your locality — the town, the district of a city, or the cluster of villages a customer would actually name — and there you are: same business, same day, same website. The only thing that changed was the size of the map you drew around the question.

    By locality we mean the local area a customer would actually say out loud — usually a town, but just as often a named district of a larger city or conurbation, or a group of small, close-together villages that people treat as one place. It is the level below the county or region, and above the single street.

    If you are going to check whether AI recommends you, this is the first thing to get right. Get it wrong and you will either panic over a problem you don’t have, or miss a real one.

    What we tested

    In a dry run of our AI Visibility Check, we took a set of local businesses in one large county and asked the same kind of question three different ways: [1][2]

    • County level — “best [businesses] in [the whole county]”
    • Locality level — “best [businesses] in [the firm’s own town or district]”
    • Auto — letting the tool decide the area for itself from the business’s website

    A large county is a good test because it is big — big enough that a customer at one end is not choosing between their local firm and one an hour away at the other end. [3] So “best in the county” and “best in my town” are not the same question at all.

    What happened

    At county level, not a single one of the fourteen businesses was named. [4] Zero. Asked to name the best across the whole county, the assistant simply didn’t reach any of them.

    At locality level, more than half were named — 21 of 38 businesses in the larger sample. [5] The businesses hadn’t changed. The question had.

    And when we let the tool pick the area itself, it leaned towards the broad end: for fourteen businesses it chose the whole county six times, the actual locality only five times, and twice — for firms on the county’s edge — it decided they were in the neighbouring city instead. [6] Left to its own devices, it named under a quarter of them. [7]

    Why a county question makes you invisible

    A county is not a market. “Best in the county” asks the assistant to pick a tiny handful of standouts for an enormous area, so almost every ordinary, perfectly good business gets left off — not because it is weak, but because the question is far too wide for it to be a sensible answer. Narrow the question to the locality a customer would actually search, and the same business becomes a reasonable thing to name.

    This is why a scary-looking result can be meaningless. “We’re invisible across the whole county” is nearly true for almost everyone, and tells you very little. “We’re invisible in our own locality, where a customer would look” is the result that actually matters.

    The trap hiding in quick checks

    Here is the practical catch. Automated “AI visibility” checkers often have to guess your area, and many default to the broadest one unless your website tells them otherwise. If your site doesn’t clearly state the area you serve, a checker can end up testing you at county level — and hand you an “invisible” verdict that a locality-level question would have overturned. In our run, the tool even placed two genuine businesses in the wrong city because their location signals were ambiguous. [6]

    So an automated check is only as good as the area it picked. If nobody looked at that, the number underneath it may be answering the wrong question.

    What to do about it

    • Check yourself at the level a customer searches: your locality, not your county or region. If a tool or a report shows you a county-wide “invisible,” re-ask about your own town or district before you conclude anything.
    • State your service area in plain words on your website — the locality you serve, spelled out: the town, the part of the city, or the villages you cover. This is the single clearest signal that stops an assistant, or a checker, defaulting you to the wrong place (or the wrong city entirely).
    • Ignore county-level scores as a measure of your visibility. They are a fair test of who the region’s few headline names are — not of whether you show up where your customers are looking.

    This is exactly why our own checks fix the question to a business’s actual locality. The county view isn’t wrong; it is simply answering a different question — and it is not the one your customers ask.

    These figures come from EC’s own AI Visibility research: a single grounded AI engine, small samples, and a live snapshot of the web that changes over time. They illustrate a pattern, not a fixed market rate. Method: entityconfidence.ai/research-methodology/ · Corrections and right of reply: entityconfidence.ai/corrections-and-right-of-reply/

  • The Measure You’re Being Shown Probably Isn’t the One That Matters

    Does the person who looks after your website ever talk to you about how your website is showing up on AI searches? There’s a serious risk that the measure you are being shown is not the one that’s most important.

    The tools that do the tracking are called “citation tracking tools”, and they measure whether or not AI produces a link to your website. That’s good. Isn’t it?

    Think of the end user getting these links. They click on them, have a quick look, and move on to the next one. But they won’t remember a domain URL so they can accurately go back to it and follow through reliably. People do that by remembering brand names — that is, by remembering the name of your business. That’s what matters commercially. The number of citations (domain links) you get is not the key point. The number of brand mentions you get is. And the measure of citations can be significantly misleading as to how well you are doing.

    This mismatch is supported by independent research. Kevin Indig’s study of almost 4,000 domain appearances across four AI engines found that three-quarters of the domains received citations — source links showing in the panel — but fewer than 40% received brand-name mentions. And most importantly of all, fewer than one in seven actually had both.

    This is supported by further evidence from Semrush, which shows that fewer than one in five brands are both mentioned and cited. This isn’t accidental, and it isn’t a passing phase. It’s structural — it’s built into the way AI works.

    We have spent a year looking into this and trying to understand exactly how it works. Most people assume that if your domain name turns up in a response to a user’s question, then the brand name follows. As you’ve just seen from the evidence, that is not the case.

    The opposite is more true, however. If your brand is mentioned in a response, then the chance of the domain name turning up as well can be better than 50%. If your brand isn’t mentioned, that chance falls to around one in ten.

    The reason is rooted in the way AI works. The AI doesn’t start from your content and work outward. It starts from its training knowledge — a picture of which brands are associated with which categories, built from the whole of the web before the model was switched on. That “training knowledge” layer decides which brands to recommend. A second step then goes looking for source material to cite alongside the answer. If your brand name was in that source material, your URL surfaces as relevant.

    An important caveat: this is a general description of how AI works. Each engine has its own way of handling it, and each also has access to up-to-date information that isn’t yet in its training knowledge. That’s why, when you run the same query several times — one after another, or on different days — you’ll likely get different answers.

    If your brand isn’t recognised in that training knowledge as a relevant player in your category, chances are that citations alone will not produce brand-name appearances. The AI may well cite your work as an anonymous source while recommending your competitors by name.

    Your content is working. Your brand isn’t. You are involved in two different battles here, and most people only see one.

    The two battles, in plain terms

    The first battle is your Reputation — what the web says about you. It’s built from everywhere your name comes up: reviews, forums, social posts, industry write-ups, the general chatter of the web. It isn’t built from your website alone. This is the evidence AI-powered search engines draw on, and the more factual, consistent, and well-organised it is, the more weight it carries.

    The second battle is your AI Visibility — whether you actually get named. When a customer asks an AI-powered search engine who the best businesses in your field are, either your name appears in the answer or it doesn’t. This is the battle that decides whether you get recommended at all.

    We measure the two separately because they move independently. A business can have a solid reputation and still go unnamed, or get named on thin evidence that won’t hold. Only measuring both tells you which one is holding you back.

    Here’s the catch. The usual tools measure the links — the citations — so that’s the number most businesses end up chasing. They pour effort into content, watch their citation count tick up, and never notice that their name still isn’t appearing in the answers customers read. Fewer than one in five brands win both battles at once — and that overlap is where all the commercial value sits.

    Why one “AI visibility score” is misleading

    There is no single AI visibility number, because the different engines behave almost like opposite personalities.

    Some engines link to sources constantly but name hardly any brands — think of the endless academic footnote, plenty of references, no names. Others do the reverse: they’ll happily recommend brands by name and rarely bother attaching a formal link. In Indig’s data these differences were stark — one major engine named brands roughly four times as often as another for the very same searches.

    So a report that says “your AI visibility is healthy” based on citations from one engine can be telling you almost nothing about what’s happening on the others. In a fifth of cases across engines, the AI systems didn’t even agree with each other on whether to name the same brand. One engine would recommend a company by name every time; another would quietly use it as an unnamed source every time. Same business, completely different treatment. That’s four separate scoreboards, not one — and a single headline figure hides it.

    The kind of content you publish decides which outcome you get

    This is the part most businesses get wrong without realising.

    Explainer and reference content — the “how does this work” material — tends to get used as an anonymous source. The AI lifts the information, answers the question, and never names you. Useful to the AI; invisible to the customer. In the research, some big reference sites were cited dozens of times and named almost never.

    Content that makes a judgement — named comparisons, recommendations, “who’s best for this” pieces, case studies where your name is attached to the result — is what actually produces brand-name mentions. When your name is inseparable from the point being made, the AI can’t use the point without saying who you are.

    The type of question matters too. The same business can show up in half of all “what’s the best…” answers and almost never in the “how does… work” answers on the exact same subject. So if the tracking you’re shown leans towards the explainer-style questions, you’re being measured on the part of AI search that matters least commercially — and missing the part where recommendations are actually made.

    What you should actually be watching

    The single most useful thing to measure isn’t your citation count. It’s the gap between how often AI uses your content and how often it actually names you. A large gap — lots of links, few mentions — is the warning sign. It means your material is good enough to borrow but your brand isn’t strong enough to recommend.

    One honest word of caution: AI doesn’t change its mind overnight. When you improve your content and your wider presence, it typically takes something like four to eight weeks for that to feed through into the answers. So this is a trend to watch over months, not a dial that moves the day after you make a change. Set your baseline now, and track the direction of travel.

    What actually closes the gap

    There are three things that move the needle, and they work together.

    Put your name inside your best claims. This works on your AI Visibility. The most immediate lever, and one you control entirely. Write your content so the insight and your business name can’t be separated — if the AI can’t make the point without naming you, it won’t. This is editorial discipline, not new infrastructure, and it’s the surest way to turn a use of your content into your name actually appearing.

    Build your presence as a recognised name in your field. This works on your Reputation. The slow, foundational work — being consistently and clearly identified across the web as who you are and what you do, so the evidence AI draws on all points the same way. It’s what drives the whole “get talked about, then get recommended” effect. It takes longest and matters most.

    Get named by others in the places AI trusts. This builds your Reputation and, in turn, your AI Visibility. Independent coverage, comparisons, recommendations and “best of” mentions from credible sources. The lion’s share of AI’s citations come from this kind of earned coverage, not from your own website — which is exactly why on-site content alone can never win it for you, and why a credible third party naming you is the strongest signal of the lot.

    Do only the first, and you’re polishing the footnote. The other two build your Reputation — the evidence AI leans on — and that Reputation is what turns into the AI Visibility that actually brings customers in.

    The bottom line

    This is why we measure two separate things: your Reputation — what the web says about you, the evidence AI draws on — and your AI Visibility — whether you actually get named in the answer. The gap between the two is the whole game, and most of the tools and reporting you’ll be shown only ever measure one side of it. They tell you whether AI finds your content useful. They don’t tell you whether AI is actually sending customers your way.

    If the report you’re looking at counts links but not mentions, it’s counting almost everything except the thing that matters. Fewer than one in seven AI appearances deliver both. Knowing which side of that line your business sits on is where real AI visibility work begins — and it’s exactly the gap we help businesses close.

    Sources: Kevin Indig’s study of 3,981 domain appearances across four AI engines; Semrush’s analysis naming the “Mention-Source Divide”; Seer Interactive’s analysis of 541,213 AI responses; Ahrefs’ analysis of AI source behaviour.

  • Identity Drift – Your AI Visibility’s Silent Killer

    A brand earns coverage in fifty articles over two years. Good sources, reputable outlets, strong editorial context. By any reasonable measure, its AI visibility should be building.

    But when you test the brand name in ChatGPT, Gemini, and Perplexity, it barely appears. When it does appear, the answers are inconsistent — right in one query, wrong in another, absent in a third.

    The coverage is real. The problem is that the AI cannot tell it is all about the same brand.

    This is identity drift — what Kurt Fischman of Growth Marshal calls the “silent killer” of AI brand visibility. Silent because no individual piece of content is wrong. Killer because the cumulative effect is invisible until you run the numbers. And it almost certainly affects your brand right now.

    What Identity Drift Is

    Identity drift occurs when a brand appears under multiple name variants across its published content and structured data, without explicit canonical links connecting those variants to a single entity.

    An example. A brand’s published content uses:

    • “Entity Confidence” (formal brand name)
    • “EC” (abbreviation, used informally in team posts)
    • “entityconfidence.ai” (domain name, used as identifier in some PR coverage)
    • “the EC team” (used in founder-authored content and podcast appearances)

    From a human reader’s perspective, these are obviously all the same thing. From an AI retrieval system’s perspective, each unlinked variant is a candidate for a separate entity. Evidence accumulates under four separate nodes. None achieves the recognition threshold that the consolidated entity would have reached.

    The error is structural, not factual. No individual page contains wrong information. The damage is done by the aggregation — fifty articles of genuine coverage, split four ways, producing four weak signals instead of one strong one.

    Why the AI Cannot Just Figure It Out

    Understanding identity drift requires understanding how AI systems resolve entity mentions — and why the process breaks down when naming is inconsistent.

    AI retrieval systems run a five-stage entity resolution pipeline:

    Stage 1 — Named Entity Recognition. Every text span is classified: is “EC” an Organisation, a Location, an abbreviation? Consistent capitalisation, explicit schema @type declarations, and contextual identifiers like “Ltd” or “Inc.” all raise the confidence of this classification.

    Stage 2 — Entity Linking. The recognised mention is mapped to a specific canonical entry in the model’s knowledge base. This is where sameAs properties in your JSON-LD do their work — they explicitly tell the model that your website entity is the same as the Wikidata entry, the same as the LinkedIn Company Page, the same as the Crunchbase profile. Without these links, the model is working from text alone.

    Stage 3 — Attribute Verification. Once linked, the model retrieves attributes: products, leadership, locations, capabilities. Confidence in each attribute depends on source agreement across the web. When your website, Wikidata entry, and LinkedIn page all state the same founding date and CEO name, confidence is high. Conflicting sources reduce it.

    Stage 4 — Temporal Validation. Freshness is evaluated. Stale data that conflicts with more recent signals reduces overall confidence, regardless of original accuracy.

    Stage 5 — Cross-Reference Verification. Claims are compared across multiple independent sources before recommendation. Entities with consistent, corroborated data across independent sources achieve the highest confidence scores. The output is a probability distribution, not a binary pass/fail.

    Identity drift attacks Stage 2. When a brand appears as “Entity Confidence,” “EC,” “entityconfidence.ai,” and “the EC team” without sameAs links connecting them, the entity linking step fails to consolidate evidence. For a brand with a common or abbreviated name, “EC” could be East Coast, Ecuador, or any of a dozen other organisations. Without structured data linking “EC” to the same canonical identifier as “Entity Confidence,” the model has no reliable basis for consolidation.

    The system is not being careless. It is designed to prioritise confident entity resolution. A brand with fragmented naming is, by design, deprioritised in favour of a brand that has made its identity easy to resolve.

    The Compounding Problem

    Identity drift worsens over time. A small team publishing consistently will accumulate drift across a growing content catalogue — blog posts use the abbreviation, press releases use the formal name, podcast appearances use whichever comes naturally in context.

    No individual page creates a visible problem. But the ratio of canonical to off-canonical mentions shifts steadily. At some threshold, the evidence fragmentation is severe enough to suppress recommendation frequency across all variants.

    The three states:

    Single canonical name, consistent sameAs links: evidence consolidates under one entity node; salience rises; AI recognition is reliable.

    Multiple variants without canonical linking: evidence fragments; each variant falls below the recognition threshold; AI confidence is low.

    Mixed (canonical plus some drift): partial consolidation; recommendation frequency varies by query phrasing; visibility is inconsistent and difficult to diagnose.

    Most brands with an identity drift problem are in the mixed state. They can be recommended — and sometimes are — but the inconsistency makes it look like a platform-specific or query-specific problem when it is actually a naming problem.

    Conflation: Worse Than Invisibility

    Incomplete disambiguation produces a second failure mode beyond low recommendation frequency: conflation.

    Conflation is when the AI resolves a name to the wrong entity — a different company with a similar name, a subsidiary with an overlapping product line, an acquired brand that no longer exists independently. The model is not refusing to answer. It is answering confidently with information about a different brand.

    This is harder to detect than absence. A brand that is absent from AI answers generates a clear gap. A brand that is conflated with a competitor generates confident, plausible-sounding wrong answers — wrong founding date, wrong product description, wrong market position — that most users and many brand teams will not notice unless they are specifically auditing AI output.

    The common triggers: generic or widely shared brand names; trading name used publicly alongside a different legal entity name; acquired companies whose independent web presence still ranks well; regional subsidiaries operating under slightly different identifiers. Each inconsistency raises the probability that the model will hedge or conflate rather than identify it cleanly.

    The Audit

    Four steps to establish whether identity drift is suppressing your brand’s AI visibility:

    Step 1 — List all name variants in active use. Audit published pages, structured data, social profiles, PR coverage, and team-authored content for every form in which the brand is referenced. Include abbreviations, domain name references, legacy names, and product variants. The list is typically longer than expected. Any variant appearing in third-party coverage without a canonical sameAs link connecting it to your registered entity is a fragmentation risk.

    Step 2 — Check sameAs coverage in JSON-LD. Every page that includes Organisation schema should have sameAs links pointing to the same canonical external identifiers: Wikidata entity URL, LinkedIn Company Page, Wikipedia article (if applicable), Companies House or equivalent, Crunchbase profile. Inconsistency across pages introduces its own disambiguation noise. Missing sameAs links leave the model without structured evidence to resolve variants.

    Step 3 — Run Google NLP API entity extraction on key pages. The API returns exact salience scores and entity classifications for the content it analyses and uses the same NLP infrastructure as Google Search and AI Overviews. A well-optimised page should show 2–3 primary entities with scores above 0.5, supporting entities clustered at 0.2–0.4, and unrelated entities absent. If the brand’s canonical name is returning a salience score below 0.5 on its own About page, that is a measurable identity problem. If abbreviations or domain variants are returning as separate entities, drift is confirmed.

    Step 4 — Test canonical and variant names in major AI engines. Query ChatGPT, Gemini, and Perplexity with both the canonical name and common variants. Compare the responses: is representation consistent? Does one variant produce different — or absent — answers compared to another? Inconsistency between query variants points to fragmented entity evidence.

    The Fix

    Establish one canonical name and use it consistently. Choose the form that appears most clearly in official registration, structured data, and high-authority third-party sources. Define it formally — a one-line canonical entity statement — and use it as the reference for all new content.

    Add sameAs links to all entity declarations. Every page with Organisation schema should link to the same canonical external identifiers. Person schema for founders and key executives should link to LinkedIn profiles with worksFor connecting back to the Organisation. The model uses these relational signals to build a structured graph of who you are and who is associated with you.

    Audit and prioritise existing content. Full remediation of every off-canonical reference in an existing content archive is rarely practical immediately. Prioritise the highest-traffic pages, the most-linked-to pages (check backlink data), and pages that rank for queries where you want AI recommendation. Update their canonical name usage and structured data first.

    Build an entity registry. A one-page internal document per entity — canonical name, description, Schema.org type, official identifiers (Wikidata QID, Wikipedia URL, LinkedIn URL, company registration number) — that gates all new content publication. Before any new page is published, the entity references on it should be checked against the registry. This prevents drift from accumulating in the first place.

    The evidence that this works is concrete. Schema App found that scaling entity linking through consistent sameAs properties and NAP data produced a 46% increase in impressions and a 42% increase in clicks for non-branded queries — direct evidence that disambiguation signals have measurable reach effects, not just AI-recommendation effects.

    Building Salience Once Drift Is Fixed

    Eliminating identity drift is the prerequisite. Once evidence is consolidating under a single canonical entity, the next step is building the salience of that entity in content.

    Entity salience — the NLP score (0–1) measuring how central a named entity is to a document — is not the same as mention frequency. Repeating a brand name ten times without a clear semantic frame produces lower salience than using it once in the subject position of a strong opening sentence.

    Growth Marshal’s Kurt Fischman describes a four-stage operational sequence:

    Stage 1 — Define canonical entities. Create a one-page brief per entity: canonical name, description, Schema.org type, key properties, and official identifiers (Wikidata QID, Wikipedia URL, LinkedIn URL). Keep URLs stable. Gate all new content publication on this registry.

    Stage 2 — Rewrite core pages. Lead with identity, explain the role, show relationships. Keep paragraphs tight. Remove metaphors that introduce ambiguity. Each paragraph should function as a semantic capsule: a subject-verb-object sentence that pins the entity to an attribute, maintained proximity between entity and defining content, a restatement at the close. Every sentence should advance the model’s understanding of which entity matters and why.

    Stage 3 — Mirror the narrative in JSON-LD. Use Organization and Person for the firm and leadership; Product or Service for offerings; CreativeWork for papers and guides; DefinedTerm for proprietary concepts. Link sameAs to official profiles. The markup should translate what the prose already says — not invent a different story.

    Stage 4 — Measure and iterate. Run entity extraction monthly via the Google NLP API. Track salience ranks and spreads: the primary entity should rank first by a wide margin; secondary entities should cluster below; unrelated entities should not appear at all. Test prompts in major AI assistants to observe recommendation lift.

    The signals with the highest individual impact: subject position (“Entity Confidence is…” assigns maximum salience; “the concept introduced by…” does not); H1 prominence (the heading is the strongest document-level salience signal); and first-100-word placement (entities introduced in the opening paragraph receive higher salience than those introduced later).

    The Sequencing Rule

    Lumar’s GEO framework makes a structural point that applies directly here: entity GEO must precede brand authority GEO.

    Before an AI platform can decide whether to trust or recommend a brand, it must first know who the brand is — which specific organisation, product, or person the content refers to. Brand authority signals, earned media coverage, E-E-A-T indicators — these have nothing to attach to if entity recognition has not been established. They accumulate under weak, fragmented nodes and fail to translate into recommendation.

    The ordering matters because it determines which work generates compounding returns. Identity clarity is the prerequisite layer. Built correctly, it makes everything else — earned coverage, structured content, cross-platform consistency — contribute to the same entity node. Built incorrectly, or left unbuilt, it causes everything else to work against a fragmented target.

    This is why identity drift is the first audit, not the last. Not because it is the most sophisticated problem, but because every other investment in AI visibility is contingent on solving it first. The brand that fixes its entity layer before adding more content will compound every piece of coverage it has already earned. The brand that skips this step will keep publishing into a leaking bucket.

    Fix the container before you fill it.

  • Writer-Only Mode: The Query That Never Reaches Your Page

    Every AI search optimisation strategy rests on an assumption so basic it is never stated: your content has a chance of being retrieved.

    For a significant category of queries, that assumption is wrong.

    The Architecture Nobody Is Talking About

    In March 2026, Baidu Search published an engineering paper co-authored by 24 practitioners documenting the internal architecture of a production AI search system running at industrial scale. This is not a research proposal or a theoretical model. It is a description of how a major AI search engine actually works, written by the engineers who built it.

    The paper describes a four-agent collaborative architecture. A Master Agent assesses every incoming query and determines which of three agent configurations handles it. If the query is complex enough to need decomposed sub-tasks, a Planner Agent maps them into a Directed Acyclic Graph — a dependency-ordered sequence where some sub-tasks must complete before others begin. An Executor Agent runs each sub-task by invoking external tools: web search, calculators, external APIs. And a Writer Agent synthesises everything into the final response.

    The paper formally adopts Anthropic’s Model Context Protocol (MCP) as the standard tool interface for the Planner-Executor pipeline. AI search at production scale is no longer a retrieval system with a language model attached. It is a reasoning system that invokes retrieval as one of several possible tools.

    What makes this architecture consequential for brand strategy is what happens before any of it kicks in. The Master Agent’s first job is not to retrieve anything. It is to decide whether retrieval is necessary at all.

    The Decision Before Retrieval

    Every incoming query is classified by the Master Agent and routed into one of three configurations:

    Writer-Only: Simple queries answered entirely from the model’s internalised training knowledge. No web search. No retrieval. No external data fetched. The model writes the answer from what it already knows.

    Executor-Inclusive: Moderately complex queries requiring one piece of external information. A single retrieval step runs, then the Writer synthesises the result.

    Planner-Enhanced: Complex queries decomposed into a directed graph of sub-tasks, each assigned a specific tool, executed in dependency order.

    The paper gives an example of a Writer-Only query: “What is the name of Emperor Han-Wu?” The answer is in training data. Retrieval adds nothing, so no retrieval occurs. (source: Towards AI Search Paradigm.md)

    The example is deliberately simple. But the principle extends to a much larger class of queries: any question where the Master Agent assesses that the model’s training knowledge is sufficient to generate a confident, accurate response. For many everyday brand queries — what a company does, where it sits in a market category, what it is known for — that assessment is made before a single URL is fetched.

    What Writer-Only Actually Means for Brands

    The two-stage citation pipeline — retrieve a pool of URLs, evaluate candidates, open selected pages, cite — applies only to queries routed to Executor or Planner mode. Writer-Only queries never reach the pipeline.

    No URL is fetched. No content is read. No website is visited. If your brand is relevant to the answer, the only way it appears in the response is if it is already embedded in the model’s training knowledge.

    This is not a fringe category. Simple factual queries — product categories, brand comparisons, general “what is X” questions — are precisely the everyday queries where brands most want to be represented accurately. They are the queries buyers ask early in the decision process, before they move to the specific comparison and evaluation questions that trigger multi-step retrieval.

    The implication is blunt: a significant fraction of the queries that most matter to a brand are answered before any retrieval occurs. Content strategy, earned media, search rankings — none of it matters for those queries. The training-data channel is the only channel that exists for them.

    The Evidence for How Deep the Bias Runs

    Researchers at the University of Toronto tested this empirically. In a 2026 study of 1,000 queries across five AI systems, they ran perturbation experiments on brand ranking queries: randomising retrieved evidence, restricting the model to only provided snippets, swapping entity names across documents. For well-known brands — major automotive, consumer electronics, household names — the rankings barely moved.

    The Kendall τ correlation between normal and perturbed rankings for popular entities was 0.911. Near-perfect consistency, regardless of what was done to the retrieved evidence. The model’s internal representation of those entities is so stable that retrieval barely influences it. (source: Navigating the Shift A Comparative Analysis of Web Search and Generative AI Response Generation.txt)

    For niche entities, the same perturbations caused rank deviation to roughly double. Strip the retrieved evidence, and niche entity rankings become unpredictable. The model has no stable prior — it follows wherever the retrieval leads.

    The citation miss rates make the same point in a different way. For popular brands in the automotive category, the study logged how often a brand appeared in an AI response without any supporting retrieved snippet:

    • Toyota: 6% citation miss rate
    • Honda: 3%
    • Kia: 10%
    • Chevrolet: 26%
    • Cadillac: 58%
    • Infiniti: 73%

    For Infiniti and Cadillac, the majority of ranked appearances came without any retrieved evidence. The model is drawing on training priors and presenting the output with the same confidence as a fully evidenced response. (source: Navigating the Shift A Comparative Analysis of Web Search and Generative AI Response Generation.txt)

    The AI’s position on Cadillac and Infiniti is not based on what any website currently says. It is based on whatever snapshot of those brands was encoded in training. If that snapshot is outdated or incomplete, no amount of current content investment will correct it until the next training run.

    How AI Learns Without Citing: The Reddit Paradox

    Understanding why training-mode representation is so stable requires understanding where it comes from. And here the data from the citation channel reveals something counterintuitive.

    ChatGPT draws URLs from five distinct retrieval channels when it does browse the web. The citation rates by channel:

    • Search: 88.46%
    • News: 12.01%
    • Reddit: 1.93%
    • YouTube: 0.51%
    • Academia: 0.40%

    Reddit is pulled at massive volume — over 16 million data points in one study of 1.4 million prompts — and cited at a rate of under 2%. The research team concluded: ChatGPT uses Reddit extensively to understand topics, gauge consensus, and build background context, then cites more institutional sources when formulating its response. The phrase they used: “It learns from the crowd, then cites another institution.” (source: Why ChatGPT Cites One Page Over Another (Study of 1.4M Prompts).md)

    This is the mechanism behind training-mode knowledge. The sources that shape what a model knows about a brand are not the same as the sources it cites. Community discussion — Reddit threads, forum posts, Quora answers, industry discussions — builds the model’s tacit understanding of what a brand represents and where it sits in a category. That understanding then governs the Writer-Only response.

    A brand that exists primarily in press releases and well-structured product pages may have excellent citation-channel signals and almost no community-level signal. For Executor-mode queries, it may be cited correctly. For Writer-Only queries, the model has little to draw on.

    The Adversarial Layer: Why Inconsistency Is Penalised

    The Baidu paper describes a further architectural decision relevant to brands with fragmented training-mode presence. The system includes adversarial robustness training: a separate agent deliberately introduces noise, fabricated content, and misleading documents into retrieval sets during training — and the Generator is trained to remain accurate despite these attacks.

    The effect is designed to protect against misinformation. The side-effect is structural: the robustness layer penalises inconsistency. A brand whose training-data profile is fragmented — different descriptions on different platforms, inconsistent entity naming, contradictory claims across sources — is more likely to be treated as unreliable noise by a system trained to reject exactly that pattern.

    Adversarial robustness and brand consistency are the same problem, viewed from different angles. The AI system is not being selective about your brand in particular. It is applying a generalised filter for signal quality. Brands that pass the filter — consistent entity naming, corroborated claims, cross-platform coherence — are treated as reliable. Those that do not are filtered out, regardless of the volume of content that exists about them. (source: Towards AI Search Paradigm.md)

    The Two-Channel Strategy

    Most brand AI visibility work is focused on the citation channel: earning coverage in sources AI engines retrieve, ensuring pages are parsable and semantically structured, maintaining search rankings so content enters retrieval pools.

    That work is necessary for queries that trigger Executor or Planner mode. For Writer-Only queries, it does nothing.

    The practical split:

    Citation-channel queries (moderately complex, multi-step, time-sensitive, comparison-focused): invest in fresh earned coverage, structured extractable content, search rankings, consistent publication in trusted sources. These queries trigger retrieval. What you publish today can influence the response in the next query cycle.

    Knowledge-channel queries (simple factual, category-level, definitional, early-funnel): invest in Wikipedia presence, Wikidata records, consistent entity naming across platforms, long-cycle community discussion, and persistent authoritative mentions that will enter future training runs. What you do today has effects measured in months.

    The two channels are not alternatives. They serve different populations of queries. A brand invested in only one is invisible in the other.

    The Timing Asymmetry

    The timing problem is where most brands get into trouble, because it is counterintuitive.

    Citation-channel work can have visible effects within days. A new piece of coverage in a trusted outlet can enter the retrieval pool in the next query cycle. The feedback loop is short.

    Knowledge-channel work compounds slowly. Training data has a cutoff. Future training runs will incorporate new information, but the cycle is measured in months to years, not days. Content published today does not update the model’s training-mode knowledge today.

    This creates an asymmetric urgency. When a brand realises that Writer-Only queries are generating wrong or incomplete answers, the correction cycle is already months away. The brands with accurate, stable training-data representation built it continuously — not reactively, not in response to a problem that had already emerged.

    The practical implication: knowledge-channel investment should start before you need it. If a brand is not actively building its training-mode presence, it is not holding steady. It is falling behind as models continue to train on an information environment the brand is not actively shaping.

    The Invisible Query

    The question for any brand is: what percentage of the queries relevant to your category are answered in Writer-Only mode?

    There is no direct way to measure this from outside the system. But the architecture tells you where to look. Queries that are simple, factual, and stable — definitional questions, brand category placement, market positioning — are the strongest candidates. They are the questions buyers ask early in the decision process, before they reach the specific comparison and evaluation queries that trigger retrieval.

    The query that never reaches your page is not the rarest kind. It may be the most common kind. And for most brands, it is the one nobody is optimising for.

    Date: 2026-09-03