Generative Engine Optimization: Six Popular Statistics, Traced to Source

Open ten guides to generative engine optimisation and you meet the same four figures: a 40 per cent lift, an 85 per cent share, a 2.8 times multiplier, a threefold penalty for stale pages. Nobody cites a sample size. We went looking for the original pages behind six of the most-repeated numbers in this category.

This is written for the person assembling an AI search plan who has to defend the slide. Where a primary document existed, we read it. Where only a vendor summary existed, we say so and mark the claim unverified rather than guessing at a method. We sell content software, so we have an interest in this subject; no competitor's product is under test here.

Six numbers, traced to whatever stands behind them

The number	Traced to	Our read	Why
A 40 per cent visibility lift from GEO	Aggarwal and co-authors, KDD 2024	Partly holds	Real peer-reviewed finding, measured on a query benchmark
AI Overviews need their own optimisation layer	Google Search Central	Does not hold	Google's own documentation says the opposite
85 per cent of brand mentions begin off your site	AirOps, echoing Muck Rack	Unverified	Direction is well supported; the figure moves between editions
Structured pages collect 2.8 times the citations	AirOps	Unverified	No method published alongside the multiplier
Skipping a quarterly refresh triples your losses	AirOps	Unverified	Two different figures, two different windows, one vendor
AI summaries suppress outbound clicks	Pew Research Center	Holds	Tracked browsing behaviour, stated sample, stated month

The 40 per cent arrived from a benchmark, not from a website

Traced to: a paper called GEO: Generative Engine Optimization, written by Pranjal Aggarwal with five co-authors, posted in November 2023 and accepted at KDD 2024. It is the document that put the phrase into circulation at all.

What the paper states: the authors describe a black-box optimisation framework plus an evaluation set they call GEO-bench, and report that their methods can raise source visibility in generative answers by up to 40 per cent. They also note that how well it works depends on the domain being queried.

Our read: partly holds. The finding is real and the authors are careful. What travels badly is the framing. A ceiling figure from a controlled query set is not a forecast for one company's blog, and the abstract never publishes the baseline the lift is measured against.

How to use it: treat the levers the paper tested, denser statistics, named citations, direct quotation, as experiments worth running on your own pages. Treat the headline figure as the upper bound of a laboratory result.

Google documented the absence of a secret layer

Traced to: nothing, which is the point. The claim that AI answers require a parallel discipline is implied by guides rather than sourced by them.

What the page states: Google's documentation on AI features and your website says there are no additional requirements and no special optimisations needed to show up in AI Overviews or AI Mode, no bespoke schema.org type to add, and that the advice is to keep important content present as text.

Our read: does not hold, at least for Google's surfaces, on Google's own account. Other assistants publish far less, which is where the uncertainty genuinely lives.

How to use it: confirm indexing and snippet eligibility before you buy anything. For the assistants that do not document themselves, our breakdown of what a page looks like to OpenAI's crawlers is the closer substitute.

The off-site share: solid direction, unstable digit

Traced to: AirOps, which carries the figure in its citation tracking guide and its blockers post, stating that roughly 85 per cent of brand mentions in AI answers originate on third-party pages. No sample is published, and the same figure reappears elsewhere on the site labelled as top-of-funnel visibility, which is a different quantity.

What newer work states: the underlying direction holds up well. Muck Rack's Generative Pulse project publishes a recurring study under the title What is AI reading?, with a May 2026 edition that describes analysing millions of consumer prompts and every source the assistants cited. Worth noting for a claim check: the landing page states the conclusion in prose and carries the share itself inside a chart image, so the number you see quoted in press coverage is not available as text on the publisher's own page.

Independent corroboration: a 2025 paper from Nick Koudas and three co-authors at Toronto, Generative Engine Optimization: How to Dominate AI Search, compared how ChatGPT, Perplexity and Gemini select sources against Google's selection across several verticals and languages. They report a systematic and overwhelming preference for earned media over brand-owned and social content, and they publish no percentage at all. That is the honest shape of this evidence.

Our read: unverified as a number, well supported as a pattern. Build the plan on the pattern.

A multiplier is not a method

Traced to: an AirOps post on content budgets, which says pages carrying sequential headings and rich schema collect 2.8 times the AI citations that unstructured pages do. It links onward to an AirOps report and does not describe how the comparison was constructed.

What the same vendor states elsewhere: its blockers post is noticeably more cautious, warning against treating markup as a rescue and offering a smaller effect, roughly a tenth higher citation likelihood for pages carrying three or more schema types. Two numbers, one publisher, no visible method behind either.

Our read: unverified. Clean heading order and accurate markup are worth doing because they make a page legible to readers and parsers alike. Whether they cause a citation is not demonstrated by either figure.

Quarterly, or annually? The rule shifts between pages

Traced to: AirOps research from 2026, restated in its tracking guide, which holds that pages left unrefreshed for a quarter shed citations three times as fast. Neither the sample nor the observation window appears.

What the companion post states: the blockers article sets the threshold at a year rather than a quarter and describes the effect as roughly a doubling. A rule that moves by a factor of four depending on which page you land on is a hypothesis wearing a number.

Our read: unverified. Recency plausibly matters; the specific cadence is invented. Refresh what has decayed or gone factually stale instead, which you can find from your own data with our year-over-year Search Console comparison.

One figure came with its homework attached

Traced to: the Pew Research Center, July 2025, working from tracked browsing rather than a survey panel's recollection.

What the study states: across 900 US adults observed in March 2025, a visit to a results page containing an AI summary ended in a click on a conventional result in 8 per cent of cases, against 15 per cent where no summary appeared. Clicks on links inside the summary itself were rare.

Our read: holds, with its boundaries stated. One country, one engine, one month in early 2025. Quote it with those boundaries and it stays useful; strip them and it becomes another floating statistic.

What separated the two groups

The numbers that survived tracing all arrived with three things: a named sample, a date, and a description of how the measurement was taken. The ones that dissolved arrived with a logo and a decimal point. The decimal is the tell. A real effect measured on a real sample gets reported with its uncertainty; a number manufactured to make a point gets reported to one decimal place and never revised.

The category is being priced before it is being proven

Here is the position we will defend: the weakest statistics in this field are not accidents, they are collateral from a land grab, and marketers should read an unsourced GEO multiplier as a pricing signal rather than a research finding.

The money makes the incentive legible. Wikipedia's entry on the discipline records, with dated reporting behind each one, Adobe's planned purchase of Semrush at 1.9 billion dollars in November 2025, Profound raising 96 million dollars in February 2026, and Sitecore reported as acquiring Scrunch for 225 million dollars in June 2026. Capital at that scale needs a category with a name, and a named category needs numbers that imply a gap classic search cannot close.

Forrester's Nikhil Lai put the structural version of this plainly in 2025, arguing that answer engine optimisation and its relatives are "significantly, but not fundamentally, different from SEO" and that advocates inflate the distinction to carve out what he called a startup-sized hole in the marketing stack. Google's own 2026 guidance lands in the same place: optimising for generative search is optimising for search.

Of course, if the unsourced figures turned out to be directionally right, this check would lose its point. That is testable, and it is the condition we are watching: if a vendor publishes its sample and its method and the 2.8 times multiplier survives, the number was simply early rather than invented. Until a method appears, the asymmetry decides it. A traceable claim costs a vendor nothing to document, so a claim that stays undocumented after a year of repetition is usually undocumentable.

Note what this position does not say. It does not say GEO work is wasted. Earning off-site mentions, writing pages that answer a question in text, and keeping facts current are all defensible on evidence we traced above. The advice is mostly fine. The arithmetic attached to it is marketing.

Trace any statistic in ten minutes

1. Follow the citation to its own citation, and keep going until a page either runs a study or admits it is quoting one.
2. Write down four fields before you read the finding: who measured, what they measured, how many, and when.
3. Search the figure as a phrase and check whether every page defines it identically. A number that changes meaning between pages was never one number.
4. Ask what the publisher sells, and whether the claim would appear if the answer were small.
5. If no method is published, record it as unverified and move on. That is a complete answer, not a failure to find one.

Run that loop and most of this category's statistics collapse into two or three findings you can actually stand behind. Those are enough to plan with. If you want the measurement side instead of the claim side, our spreadsheet playbook for AI citations builds the baseline from your own observations, which is the only dataset nobody can misquote at you.

Questions marketers keep asking about GEO figures