At a glance
- AI engines do not cite pages, they cite claims, and a claim becomes citable only when it is narrow enough for an engine to check.
- Across 15,000 prompts, only 12 per cent of the URLs cited by ChatGPT, Gemini and Copilot appeared in Google’s top ten for the same prompt (Ahrefs, August 2025). Ranking is no longer the qualifying test.
- Vantage has tracked twenty buyer-intent queries weekly since July 2026 across thirteen of its pillar pages. Two win their query outright. Eleven are invisible.
- The two do not have better schema, structure or age than the eleven. They make a narrower claim.
- Narrowing a claim is a positioning decision, not a technical one, which is why tooling does not fix it.
A citable claim is one an engine can verify without leaving the page
A citable claim is a statement specific enough that a generative engine can match it to a user’s question and quote it without seeking corroboration elsewhere. It names something: a sector, a regulation, a scheme, a number, a market. The test is not whether the sentence is true, since almost all brand copy is true in the trivial sense. The test is whether it is falsifiable, because a claim that cannot be wrong carries no information a retrieval system can act on.
This is not the same as being well optimised. Schema markup, clean heading hierarchy, llms.txt, freshness and passage-level anchoring all matter, and Vantage has written about each, from the groundwork for getting mentioned by ChatGPT to measuring visibility once you are in the answer. They govern whether an engine can read a page, not whether it has anything worth quoting. The distinction shows up in an awkward observation: two pages on the same domain, with identical schema, structure and age, routinely differ completely in whether an engine will cite them.
The nearest analogy is expert witness testimony. A witness who tells the court the defendant was careless has said something true and entirely unusable. A witness who says the brake pads measured 1.2mm against a 3mm service limit gets quoted verbatim in the judgment. Same expert, same knowledge, same case. The difference is admissibility, and admissibility is a property of the statement rather than the speaker.
Only 12 per cent of the URLs AI engines cite appear in Google’s top ten, so ranking is no longer the qualifying test
The assumption underneath most brand content strategy is that citation follows ranking. Get to page one and the engines will find you. The measured overlap does not support it.
Ahrefs analysed 15,000 long-tail prompts through Google, Bing and four AI assistants. On average, only 12 per cent of the links cited by ChatGPT, Gemini and Copilot appeared in Google’s top ten for the same prompt, and more than 80 per cent came from pages that did not rank at all for the target query. Perplexity was the outlier at 28.6 per cent, which makes sense given it was built to cite (Ahrefs, published August 2025, updated May 2026).
Google’s AI Overviews were long the exception, tracking organic results closely. That has loosened. An updated Ahrefs analysis of 863,000 keywords and 4 million AI Overview URLs found 38 per cent of cited pages also ranked in the top ten, against 76 per cent in the July 2025 version of the study. The caveat belongs here rather than in a footnote: Ahrefs states its parsing improved between the two, so the datasets are not strictly comparable, and a separate BrightEdge analysis put the overlap nearer 17 per cent (Search Engine Journal, March 2026). The figure is contested. The direction is not.
The mechanism is query fan-out. An engine splits one question into several sub-questions, retrieves against all of them, and favours pages that surface consistently across the set. A page ranking sixth for three adjacent sub-questions beats a page ranking first for one. That is a structural preference for content answering a specific question several ways, and against content that gestures at a category.
Ranking answers the question of whether a page can be found. Citation answers a different question entirely: whether anything on it can be quoted.
We tested this on our own thirteen pillars, and eleven of them were invisible
Since July 2026, Vantage has tracked a fixed set of twenty buyer-intent queries weekly across ChatGPT, Perplexity and Google’s AI surfaces, checking whether any Vantage page is cited in the answer. The set covers thirteen pillar pages, all built to the same template, all carrying the same schema stack, all published within a fourteen-month window. The method follows the prompt-set discipline set out in our guide to measuring brand visibility in AI search.
As at 29 August 2026, two of those thirteen win their query outright. Eleven are invisible.
The number is worth stating plainly rather than softening. It is a small sample, it is our own data, and it is not a controlled experiment. It is also the most useful thing we have, because it holds constant nearly everything the standard advice tells you to fix. Same template, same schema, same internal linking, same cadence. Whatever separates the two from the eleven is not on that list.
Any consultancy can publish a checklist. Publishing the finding that most of your own content failed it is less comfortable and considerably more informative, because it is the only version that could have come from doing the work rather than reading about it.
Update, 5 September 2026. The following week’s tracking moved the two, and the movement is worth recording rather than quietly leaving the figure to age. The grants pillar lost its top position to a competitor who published a more specific page on the same scheme, having held that position for four consecutive weeks. The positioning framework pillar appeared for the first time in this series, on the strength of naming a four-part framework rather than describing the category. One left, one arrived, so the count holds at two, and the mechanism holds with it. A specific claim wins the position and a more specific claim takes it away. Nothing about the template, the schema or the age of the pages changed in either direction.
The two that worked shared one property, and it was not schema, structure or age
Set the two winning pages beside the eleven and one difference survives inspection. The winners make a claim an engine can check. The eleven make claims about a category.
One of the two is our healthcare branding pillar, which took the top position on its query for the first time in late August 2026, corroborated independently on Perplexity and ChatGPT. It does not claim expertise in healthcare branding in the abstract. It works from the regulatory constraint Singapore providers operate under, and answers what a brand can be built from once that constraint applies. The second winner is scoped just as tightly, to a named scheme with published eligibility rules.
The eleven are not bad pages. Several are the most thorough things we have written. They are wide. A comprehensive guide to brand positioning is a claim about a category, and every consultancy makes the same one. When an engine fans a query out and looks for passages that distinguish one source from another, category claims collapse into each other. There is nothing to choose between them, so nothing gets chosen.
Exhibit 1: The specificity threshold
A single axis running from category claim to checkable claim, with a threshold partway along. Below the line an engine has nothing to verify. Above it the claim contains a fact an engine can match to a question. Each rung describes the same firm with the same capabilities. Only the width of the assertion changes.
| Rung | The claim | Position |
|---|---|---|
| 1 | “We are a leading branding agency” | Below. Nothing to verify. A superlative is not a fact. |
| 2 | “We do healthcare branding” | Below. Category membership, not a claim. Shared with every competitor. |
| 3 | “We brand regulated healthcare providers in Singapore” | At the line. Verifiable in principle. Named sector, named market, named condition. |
| 4 | “We build brands for HCSA-licensed providers who cannot use testimonials or superlatives” | Above. Contains a checkable fact an engine can match to a real question. |
The point of the exhibit is that rung four is not a smaller firm than rung one. It is the same firm, described in a way that gives a retrieval system something to hold. The narrowing happened in the sentence, not in the business.
The specificity threshold is where a category claim becomes a checkable claim
The threshold sits where a claim stops describing what kind of firm you are and starts describing a condition someone could test. Three properties appear together at that point.
It names a constraint rather than an aspiration. “We understand regulated industries” is an aspiration. “Providers licensed under a named Act may not publish testimonials” is a constraint, and a reader who works under it recognises it immediately. Constraints are checkable because they exist independently of the person asserting them.
It carries a number, a date or a proper noun, not as decoration but as the load-bearing part of the sentence. A named statute, a commencement date, a fee range, a measured result. These are what a fan-out query matches against, because users ask questions containing them.
It would be false for most firms in the category. This is the hardest property to accept and the most reliable test. If a competitor could paste your sentence onto their own site without it becoming untrue, it does not distinguish you, and a retrieval system has no reason to prefer your version.
Exhibit 2: What the two cited pillars had that the eleven did not
Drawn from the Vantage tracking set as at 29 August 2026. Observed properties, not a ranked model.
| Property | The two cited pillars | The eleven uncited pillars |
|---|---|---|
| Short answer at the top | Present. Question answered in the opening passage, extractable on its own. | Present in template, but answers a category question rather than a specific one. |
| Dense extractable numerics | Present. Named figures, dates and thresholds through the body. | Sparse. Figures appear as illustration rather than as the claim. |
| Claim scoped to a named sector or scheme | Present in both. | Absent. Scoped to a discipline instead. |
| Anchor-linked contents | Present. Passage-level addressability for fan-out retrieval. | Present in some. Not sufficient on its own. |
The fourth row matters for what it rules out. Anchor-linked contents appear on both sides of the table, which is precisely why they cannot be the explanation.
Retrieval is moving from open-web search to direct domain lookup, which raises the cost of a vague page
A change measured in August 2026 sharpens the argument. Promptwatch, which monitors the sub-queries ChatGPT generates when decomposing a prompt, recorded the share of fan-out queries using the site: operator jumping from about 0.37 per cent to 16.8 per cent in a single day on 8 August 2026, roughly a 46-fold increase. Searches run per response nearly doubled at the same time, from about 1.08 to 1.83, indicating domain-scoped searches were added on top of the general ones rather than replacing them.
Over the same window, Reddit’s share of ChatGPT Search citations fell from an average of 3.83 per cent between 18 July and 7 August 2026 to 0.52 per cent between 14 and 17 August, an 86.4 per cent relative drop (Promptwatch, August 2026, reported by Forbes, 20 August 2026). Promptwatch presents the timing as a plausible mechanism rather than a proven cause. Google’s AI surfaces declined only gradually over the same window, around 11 per cent for AI Overviews and 30 per cent for AI Mode, so this was a shift in ChatGPT’s retrieval behaviour rather than an industry judgement about forums.
The implication for brands is direct. An engine that asks named domains for answers has already decided which domains plausibly hold them, and it decides on the strength of what your domain is known to be about. Our analysis of how AI engines decide which brands to recommend covers what builds that reputation. A site whose claims are wide never becomes the obvious place to ask.
How to narrow a claim without narrowing the business
The objection here is commercial, and it is fair. A narrower claim sounds like a smaller market. In practice the reverse holds, for a reason worth stating precisely: the narrowing happens in what you publish, not in what you accept.
Start from the constraint your best clients work under, rather than the discipline you sell. Clients rarely arrive looking for brand strategy. They arrive unable to say something they need to say, or unable to distinguish themselves from four competitors who look identical. The constraint is what they can name, and what they type into an engine.
Write one page per constraint rather than one per service. A comprehensive guide to positioning competes with every other comprehensive guide to positioning. Four pages, each scoped to a sector and a named condition, compete with almost nothing, and each is retrievable against a different set of fan-out queries. This is a different discipline from traditional search optimisation, and our comparison of GEO and SEO sets out where the two diverge.
Keep the superlatives out. “Leading”, “best in class” and “trusted partner” are not merely weak, they are unhelpful to a retrieval system, because they occupy the place in a sentence where a checkable fact could have gone. This article makes no claim that Vantage has solved the problem. Eleven of thirteen tracked pillars are still invisible, which is the point.
Accept, finally, that this is positioning work rather than optimisation work. No tool can narrow a claim on your behalf, because narrowing one means deciding what you are not. That has always been the hard part of brand strategy, and it is where a brand positioning framework earns its keep. Generative engines have started pricing it.
How the specificity threshold applies in Southeast Asia
The threshold is reached sooner in a small market. Singapore’s branding category is compact enough that category claims collide almost immediately, and the collision is visible in the tracking data. On the GEO and brand visibility queries in the Vantage set, the competing pages are held almost entirely by search and digital agencies, and none argues from a positioning premise. Seven firms making broadly the same claim about the same discipline, in a market this size, produce a set of pages an engine cannot tell apart.
The regional dimension adds a second effect. Southeast Asia is not one market, so generic regional claims fail twice over, once for being wide and once for being wrong. A claim scoped to Singapore’s healthcare licensing regime does not transfer to Malaysia or Indonesia, where the instruments differ. Firms respond by writing a single pan-ASEAN page to cover everything, which is the widest possible version of the claim and reliably the least citable.
The upside for regionally focused firms is practical. Named regional specifics are undersupplied. Statutes, grant schemes, licensing regimes and commencement dates in Singapore, Malaysia, Vietnam, Thailand, Indonesia and the Philippines are exactly the checkable facts fan-out retrieval favours, and global content largely ignores them. A firm willing to write to one market’s actual constraints holds ground no volume of general brand strategy content can take.