Category: Uncategorized

Mass Adoption Is Already Here

The conversation about AI search has been framed, for two years, as a transition story. Brands were preparing for a future state. Watching adoption curves. Waiting for the numbers to become undeniable.

The numbers are now undeniable. The transition is over.

The Threshold That Just Passed

ChatGPT had 400 million weekly active users in February 2025. By July 2025 it had 700 million — growth of roughly 11–12% month over month across five months, according to OpenAI’s own published usage research.

The threshold that matters — the point Rogers’ diffusion model identifies as the mainstream tipping point — is 20% of global internet users. At 5.65 billion internet users, that is approximately 1.13 billion weekly active users.

At 5% monthly growth — the most conservative plausible rate — ChatGPT crossed that threshold around April 2026.

The window for establishing AI search presence before mass adoption normalised was narrow. It has closed. Brands that have not yet invested in AI visibility are entering the space as latecomers, not early movers.

Why This Is Different From “AI Is Growing”

Every month brings a new statistic showing AI tools gaining ground. That story is familiar and easy to discount — adoption curves always look steep in early innings.

This is not that story.

What changed in the OpenAI usage data published in September 2025 was not the scale number. It was the behaviour number.

In June 2025, asking questions — information queries of the kind that were previously the exclusive domain of search engines — accounted for 51.6% of all ChatGPT interactions. Generating content (writing, summarising, creating) had fallen to 34.6%.

ChatGPT started as a content generation tool. Within a year of that framing being established, search had become its primary use case. Not a secondary one. The primary one.

The implication: the 700 million weekly active users are not mostly doing work tasks. They are mostly asking questions that used to go to Google.

The Non-Work Signal

One data point anchors this as a consumer behaviour shift, not a productivity tool story.

In June 2024, non-work messages accounted for 53% of ChatGPT interactions. By June 2025 that figure had risen to 73%.

A tool that is 73% non-work has crossed from enterprise utility to everyday utility. It is being used to make decisions — about purchases, products, services, brands — by people who are not thinking of it as a workplace tool. They are thinking of it the way they once thought of Google: as the place you start when you want to know something.

For brands, this matters because it relocates the audience. The users who are forming opinions about your product category, your competitors, your company, are doing so through an interface that does not show them your website. It shows them whatever the model has assembled about you from the sources it trusts.

The Traffic Picture

AI-referred traffic to retail sites grew 393% year over year in Q1 2026, according to Adobe’s analysis of over a trillion visits to US retail sites.

At current penetration, that traffic is still a small fraction of total visits. But the quality signal has already flipped. In March 2025, AI-referred visitors converted 38% worse than organic search traffic. By March 2026 — one year later — they converted 42% better. They also spent 48% longer on site and browsed 13% more pages.

The implication of that reversal: being recommended by AI is already a filtering mechanism. Users who have formed a positive view of a brand through AI interaction and then choose to visit the site are arriving with higher intent than a typical search click. Volume is low now. Quality is high now. Scale is arriving fast.

The brands AI already recommends are capturing that quality audience. The brands it doesn’t are invisible to it.

What Mass Adoption Changes

Before mass adoption, a brand’s absence from AI responses was an edge-case problem. Only a small segment of users would encounter the gap.

After mass adoption, absence is mainstream exposure. A category buyer with a question — about your market, your product type, your company — is now more likely to encounter AI-mediated information than a traditional search result page.

Two shifts follow from this:

First: the strategic priority is no longer building presence ahead of a future state. It is closing a visibility deficit that already exists.

Second: the measurement framework has to change. Traffic analytics capture what a fraction of users do after they decide to visit. They do not capture what the majority of users encounter, conclude, and act on before they ever reach your site. Recommendation measurement — how often AI mentions your brand, in what context, with what confidence — is the only view into the larger surface.

The Early Mover Window

Early mover advantage in AI search follows the same logic as early mover advantage in traditional SEO. The brands that built consistent earned media presence, clear entity signals, and cross-platform consistency in 2024 and early 2025 are compounding that investment now. The knowledge base about them is richer. The recommendation rate is higher. The confidence language around them is stronger.

That window is not closed forever — AI models are continuously retrained and updated. But it does not get easier to enter a market after mass adoption than before it. The compounding that early movers have accumulated does not reset.

The question for every brand that has not yet acted is not whether to invest in AI visibility. The threshold passed in April. The question is how much gap to close, and how quickly.

*Sources: OpenAI, “How People Use ChatGPT” (September 2025), via Search Engine Journal; Adobe.

The Six Factors That Actually Decide Whether AI Mentions Your Business

AI search engines don’t rank pages the way Google does. They decide, page by page, whether a piece of content is worth quoting in an answer — and the data now shows that decision comes down to roughly the same six things, again and again, across independent studies.

The Short Version

Multiple studies — practitioner datasets covering thousands of campaigns, and academic research — converge on the same handful of factors that predict whether AI systems cite a business. None of them require a large budget. All of them require consistency.

The Six Factors

1. Can AI actually read the page? This comes before everything else. Adobe’s analysis of over a trillion US retail site visits found that product pages are, on average, only 66% readable to AI systems — homepages fare a little better at 75%. The most common cause is JavaScript-rendered content that never reaches the raw HTML an AI crawler sees. A page that fails here doesn’t get a lower score on the other five factors. It gets ignored entirely.

2. Does the content answer the question immediately? Content that states its main point in the first sentence of each section — rather than building up to it — is cited roughly 2.3 times more often than content with a traditional narrative structure, according to campaign data spanning hundreds of clients. AI systems extract answers; they don’t reward you for making them scroll to find one.

3. Is it written plainly? Content that’s needlessly hedged, padded, or complex costs citations even when the underlying facts are accurate. An analysis of how ChatGPT selects and quotes source material found that direct, definitional sentences — “X is Y” — are cited nearly twice as often as explanatory or discursive prose, and that headings framed as direct questions are twice as likely to be quoted as narrative ones. AI systems extract whichever available source states a fact most clearly — not necessarily the most authoritative-sounding one.

4. Is it current? AI-cited content skews newer than traditionally-ranked content. One large multi-platform study — 12,500+ queries across 8,000 domains — found a 3.2x citation multiplier for content refreshed within the last 30 days. Perplexity is the most aggressive platform on this factor, deprioritising unrefreshed content after just 2-3 days.

5. Are you mentioned by name, not just linked? Across several independent datasets, being mentioned by name in other people’s content correlates about 3 times more strongly with AI citation than the number of backlinks pointing at your site. A link is a reference. A named mention is a recommendation.

6. Is your business unambiguous? Name, address, and category consistency, plus basic structured data markup, form a disambiguation layer underneath everything else. Without it, the other five factors get diluted across fragmented, unrecognised versions of your brand — the AI can’t credit consistent evidence to an entity it can’t reliably identify as one thing.

What This Means in Practice

None of these six factors require a large marketing budget. They require structural discipline: checking whether your pages are actually machine-readable, writing answers before context, keeping content current, and making sure your brand is described the same way everywhere it appears. Most businesses have never audited even the first one.

The SME Advantage in AI Search

Almost every piece of advice written about AI search optimisation treats the challenge as universal. Get cited by AI. Build authority. Earn coverage. It applies to all brands equally.

The data says otherwise. A peer-reviewed study from the University of Toronto, published in January 2026, found that large, well-known brands and mid-sized niche brands are playing structurally different games in AI search — and that for one of those groups, the investment required to improve AI visibility is dramatically lower than most people assume.

Summary: The brands best positioned to move the AI visibility needle quickly are not the category giants. They are the well-regarded specialists that AI models have not yet fully formed a view about — where the next piece of coverage in the right outlet can change what AI says.

Small and Mid-sized SMEs are that group.

Read on…

Why Big Brands Are Stuck

Start with the uncomfortable finding for large brands.
Chen et al. (arXiv:2601.16858, January 2026) ran a series of perturbation experiments on GPT-4o: they manipulated the evidence the model received about a set of brands — shuffling it, restricting it, even swapping brand names into irrelevant snippets — and measured how much those manipulations changed the model’s rankings.
For well-known, popular brands, the results were clear. Changing the retrieved evidence barely moved the output.

For popular brands, even aggressive manipulation of the retrieved evidence produces low rank deviation. The model’s internal hierarchy — built during training on years of web content — dominates. The model already knows who the major players are. Retrieval is used to confirm and support what it already believes, not to discover who deserves to be recommended.

The citation miss rate data makes the same point from a different angle. For well-known automotive brands, the researchers logged how often a brand appeared in AI rankings without any supporting snippet from retrieved content:

Cadillac and Infiniti appear in AI answers without any retrieved evidence more than half the time. The model is drawing on training priors, not retrieved content. Toyota appears without evidence only 6% of the time — because mainstream outlets have covered Toyota extensively, and that coverage dominates the retrieval pool.

For a large brand trying to improve its AI positioning through content publication, this is a structural problem. The model has already formed its view. Publishing fresh content does not change a pre-trained ranking. Getting into future training runs — through Wikipedia, persistent authoritative web mentions, and community discussion at scale — is what moves the needle for large brands. That is a long-cycle activity measured in years, not campaigns.

Why SMEs Are Not Stuck

Niche brands — the mid-sized, specialist businesses that serve specific categories or audiences — face a different situation entirely.
In the same experiments, niche entities showed high rank sensitivity to the retrieved evidence. When snippets were shuffled, rankings shifted significantly. When brand names were swapped into unrelated content, the model was again highly sensitive. The critical result: when retrieval was restricted to only the provided snippets, niche entity rankings stabilised dramatically (Δavg dropped from 4.15 to 0.46).
The interpretation: the model has no stable internal hierarchy for niche brands, so it follows wherever the retrieved evidence leads. In the absence of strong training priors, retrieval is driving the answer — not confirming one.
The alignment data reinforces this. Kendall τ (alignment between holistic and pairwise rankings) for popular brands is 0.911 under normal conditions and reaches a near-perfect 1.000 under strict grounding. For niche brands, it is 0.556 under normal conditions and only 0.689 under strict grounding — reflecting genuine model uncertainty rather than stable, training-formed views.

For a mid-sized SME, this uncertainty is an opportunity. Because the model lacks pre-trained confidence about niche entities, fresh earned coverage in a trusted source can shift AI rankings in the next retrieval cycle. The investment required to change AI outcomes is proportionally smaller — and the feedback loop is measured in weeks rather than years.

The Niche Convergence Bonus

There is a further structural advantage for niche brands that the research documents.
When a query is narrow enough that both AI and Google converge on a small pool of specialist sources, the gap between SEO and AI search optimisation largely disappears. The paper found that niche queries produce 3–4 percentage points more overlap between AI citations and Google’s top-10 results than popular queries do.
For a popular brand query — “best smartphones” — GPT-4o and Google are drawing from completely different domain ecosystems (the study measured only 4% domain overlap for GPT-4o overall). For a niche query — “top ultramarathon GPS watches” — both systems converge on the same small cluster of specialist review outlets, because that is where the authoritative content lives.

This means that for mid-sized brands in specialist categories, AI search optimisation and traditional SEO are largely the same work. Securing coverage in the specialist publications that cover your category — the review sites, trade outlets, and editorial destinations your potential buyers use — builds Google ranking and AI citation simultaneously. There is no separate AI strategy required.

What This Means in Practice

The distinction between popular and niche entities implies a different investment logic for different types of brand.
For large, well-known brands, AI ranking is governed by training-time knowledge. The work that moves the needle — Wikipedia presence and Wikidata records, repeated mentions in high-authority publications over years, community discussion at scale — is the kind of brand-building that has always mattered for prominence. Short-cycle content publication produces minimal effect on AI rankings for popular queries. The horizon for this work is long.
For mid-sized SMEs, retrieval dominates. The practical priorities are:
Fresh earned coverage in trusted specialist outlets. AI systems cite content that is, on average, 62–90 days old (Chen et al.). A consistent cadence of editorial coverage in the publications that cover your category keeps your brand in the active citation pool. Historical coverage from eighteen months ago is likely no longer being cited.
Structured, extractable content. For niche entities, the model is building its answer from what retrieval surfaces. Content that is clearly attributed, well-structured, and names the brand in the evaluative sentence — not buried in surrounding context — is more likely to enter the context window and generate a recommendation.

Coverage in category-trusted outlets, not just any press. AI citation authority is concentrated in relatively small clusters of specialist publications per category. For consumer electronics, TechRadar, Tom’s Guide, RTINGS, and CNET dominate AI citations. For automotive, Consumer Reports and Car and Driver. The equivalent tier exists in virtually every category. Knowing which publications AI systems trust in your space is the highest-value strategic input for an earned media brief.

The Inversion That Most AI SEO Advice Misses

Most writing on AI search positions large enterprises as the natural leaders in this space — the brands with the biggest content budgets, the most established authority, the longest track records. In traditional SEO, that advantage compounds.
In AI search, the data suggests the opposite logic applies at the margins. Large brands are locked in by training data they cannot rapidly change. Mid-sized SMEs are operating in the retrieval layer, where fresh coverage, consistent presence, and specialist-outlet relationships translate directly into AI visibility — faster, and with a smaller investment floor.

The brands best positioned to move the AI visibility needle quickly are not the category giants. They are the well-regarded specialists that AI models have not yet fully formed a view about — where the next piece of coverage in the right outlet can change what AI says.

Sources: Chen et al. arXiv:2601.16858 (University of Toronto, January 2026); ConvertMate GEO Benchmark Study 2026.

Who Is Responsible When AI Gets Your Brand Wrong?

When a journalist writes something inaccurate about your business, you know who to call. There is an author, an editor, a publication. The correction path is clear, even if it’s difficult.

When AI says something inaccurate about your business — wrong founding date, conflated with a competitor, fabricated capabilities presented as fact — there is no one to call.

This isn’t a gap in current infrastructure that will be filled as the industry matures. It is a structural feature of how AI-generated content works, and Harvard’s Kennedy School has a name for it: distributed agency.

No one to call

The Harvard Kennedy School Misinformation Review draws a sharp distinction between two fundamentally different categories of inaccuracy.

Human misinformation is produced by actors with beliefs, motivations, and goals. The correction strategies follow from that: identify the actor, apply social or legal pressure, issue counter-statements, reduce amplification. The accountability chain exists because there is an agent at the origin.

AI hallucination emerges from a different structure entirely. A probabilistic system generates statistically likely text sequences — predicting the next token based on patterns across billions of documents. There is no intent. There is no understanding of accuracy. There is no actor who made a decision about your brand. The output is the aggregate product of training data patterns meeting a query.

Grok 3 sent 154 of 200 users to error pages in the CJR/Tow Center study. It did so with no awareness that anything had gone wrong. ChatGPT correctly identified one of ten San Francisco Chronicle articles despite Hearst’s formal content partnership with OpenAI. OpenAI did not decide to misrepresent those articles; the model generated plausible text and the text was wrong.

The practical consequence is that you cannot send a correction to a probability distribution. The remediation path for AI-generated brand misinformation is fundamentally different from correcting a journalist, a review, or a social media post. There is no actor to change their mind.

Why the belief doesn’t update

If there’s no actor to correct, you might assume the fix is at least clean: insert accurate information into the training and retrieval layers and let the model update. The reality is harder.

Mark Coeckelbergh (University of Vienna, Social Epistemology, 2025) describes the structural dynamics of AI-mediated belief as an “economy of belief revision.” Maintaining an existing belief is the cheaper option — it requires less work. Revising a belief requires more. The architecture is structurally biased toward its current representation of the world. “Using AI is believing.”

The parallel for AI systems themselves is direct. A model that has formed a representation of your brand — its authority tier, its associations, its positioning — holds that representation with inertia. Training-mode knowledge, once embedded, is not updated by a single counter-publication. It is shifted by sustained, consistent signals across multiple retrieved sources over time.

Coeckelbergh’s broader concept sharpens this further. AI-generated brand claims are not testimony in the ordinary sense — there is no human originator who holds the belief that needs revising, who can be questioned, who can update their view. They are a new epistemic category: machine-generated assertions delivered in natural language, without a human at the origin, resistant to the correction mechanisms we normally apply because those mechanisms assume an agent.

This is why correcting AI brand representation is not a one-shot content operation. It requires sustained provenance infrastructure — consistent signals across earned media, structured data, and knowledge graph entries — built and maintained across retrieval cycles to overcome structural inertia. One press release does not move the needle. One corrected Wikidata entry does not move the needle. The pattern of evidence, accumulated over time, is what shifts the model’s representation.

From answering to acting

The distributed agency problem would be serious if AI only answered questions. It becomes structurally different as AI moves from information to action.

The spectrum of agency runs from direct human browsing at one end to fully autonomous AI action at the other. The middle position — AI-assisted human activity, where software executes tasks on behalf of humans who have stated a goal — has been steadily moving toward the autonomous end.

Large Action Models are the technical substrate making this concrete. Where Large Language Models answer questions, LAMs take actions: executing task sequences, navigating interfaces, filling forms, making decisions in dynamic environments. OpenAI’s Operator and Google’s Auto Browse in Chrome can already compare products, fill forms, and make purchases without human intervention. Baidu’s production AI search system runs a four-agent architecture — a master agent classifying query complexity, a planner decomposing it into sub-tasks, an executor running tools, a writer synthesising the output — all without a human in the loop.

When AI answers a question about your brand inaccurately, a potential customer may be misled. The harm is probabilistic, downstream, mediated by human judgment.

When AI acts on behalf of a user — selecting vendors, booking suppliers, making purchase decisions based on whatever knowledge it has assembled — wrong brand knowledge doesn’t mislead. It determines concrete commercial outcomes, without any human review step between the AI’s representation and the result.

The accountability gap does not get smaller as AI becomes more capable. It gets wider, because the consequences of distributed agency are no longer hypothetical.

The trust chain

Gartner identifies trust as the primary adoption constraint for agentic AI: users will only delegate real decisions to agents they trust. But trust works in a chain.

If users trust agents, those agents must in turn trust the sources they consult. An agent making a vendor recommendation is implicitly asserting the credibility of whatever information it has assembled about that vendor. If that information is wrong — stale, conflated, fabricated — the trust chain carries the error forward, at scale, automatically.

Coeckelbergh’s concept of epistemic bubble is directly relevant here. Once a brand is positioned within a particular cluster in an AI system’s training or retrieval logic, counter-evidence may be de-weighted by the same mechanisms that created the cluster. AI-mediated knowledge environments can sustain incorrect representations not through malice but through the structural logic of how relevance and consistency are weighted. A brand that has been miscategorised, or weakly evidenced, or consistently underdescribed relative to competitors, faces an uphill correction problem that gets steeper the longer the representation persists.

The brand that is well-represented in the AI’s knowledge layer — consistent entity signals, authoritative third-party coverage, structured data that is unambiguous — sits inside the trust chain. The brand that is poorly represented, misidentified, or absent sits outside it.

When agents move from recommendation to action, being outside that trust chain is not a visibility problem. It is a market access problem.

The only available lever

None of this changes by filing a complaint. The accountability is distributed; the responses must be structural.

The Harvard HKS framework is explicit about this: the remediation path for distributed agency problems runs through the supply side. What the model has learned, and what it can retrieve, are the inputs that determine what it says and does. Those inputs are the only levers available to brands.

In practice, this means:

Earned media in publications the model retrieves reliably. The 82% of AI citations that come from editorial sources (Muck Rack, 1M+ citations) are the primary signal pool. The model learns which brands belong in which categories largely from this layer.

Structured entity data — schema markup, sameAs links, Wikidata entries — that gives the model unambiguous resolution of who you are and what you do. The technical layer that prevents conflation and suppresses drift.

Consistent, sustained signals across retrieval cycles. Not one-shot publications, but ongoing patterns of corroboration. The economy of belief revision works against you on one input; it works for you when the signals are consistent enough to shift the prior.

The question “who is responsible for what AI says about your brand?” has a clean answer: no one who can be corrected.

The question that follows is different: what is in the supply layer that shapes what the model has learned?

That is what entity confidence measures — not accountability for the current representation, but the reliability of the evidence available to AI systems when they reach for your brand. As AI moves from answering to acting, the stakes of that measurement become concrete in a way they weren’t when the only output was text.

A weak entity signal was a visibility problem when AI gave advice. When AI makes decisions, it becomes something closer to market exclusion.

Entity Confidence measures the reliability of that evidence layer.