In July 2026 a researcher at Sciences Po read every study on generative engine optimisation he could find, 45 of them, and concluded that no tactic in any of them has been shown to reliably get a website discovered by an AI engine.
That is a hard sentence to put next to most GEO sales decks, including some from UK agencies.
The paper is Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026) by Olivier Martinez. It covers work published between 16 November 2023 and 14 July 2026. This article explains what it found, in plain terms, and what that means if you run a UK business and someone is asking you to pay for GEO.
We sell this service, which is a reason to read the evidence carefully.
What generative engine optimisation is, and where the "40%" came from
Generative engine optimisation is the attempt to make your content more likely to be retrieved, cited and used when ChatGPT, Perplexity, Google AI Overviews or AI Mode answer a question. The same idea goes by several acronyms (AEO, AIO, LLMO, AI SEO), and Wikipedia's entry notes there is no settled line between them.
The field has one founding paper. Aggarwal and colleagues published "GEO: Generative Engine Optimization" in November 2023, later presented at KDD 2024. They built GEO-bench, a set of 10,000 queries, fed the top five Google results for each into GPT-3.5, and tested nine rewriting methods on one of those five sources.
Adding quotations, statistics and citations did best. Keyword stuffing scored below the unedited baseline. The headline "up to 40% more visibility" comes from one metric (position-adjusted word count) rising from 19.3 to 27.2 for quotation addition, which the survey works out as roughly 41%.
The catch, which the survey spends most of its length on, is that in that experiment the source was already in the context window. The engine was always going to read it.
The test measured how much of the answer the source got, and said nothing about whether it would have been found in the first place. Nobody clicked anything, because there was nothing to click.
The pipeline: four stages, and why most tactics only touch the last one
The survey's main contribution is to stop treating AI visibility as one ranking and model it as a chain of stages, each with its own randomness. The formal model has four.
- Activation. Does the engine search at all for this prompt? One audit in the survey found Google AI Overviews triggered on 13.7% of queries overall, rising to 64.7% for question-phrased queries. If it does not search, nothing you did to your page matters.
- Retrieval. Which candidate pages come back, given the query, the user, the engine, the date and the state of the web that day.
- Reranking. Which of those candidates get space in the context the model reads, and in what order.
- Generation. What the model writes, and which sources it cites.
The paper's own figure stretches this into a longer pipeline that adds crawling and indexing before retrieval, and absorption, fidelity and the click after generation. Limy's summary compresses it into six stages of AI search optimisation: search decision, crawl, retrieve, context, cite, absorb. The count matters less than the point. Each stage is a gate, and passing one does not pass the next.
Almost every tactic in the Aggarwal paper, and almost every tactic in the trade press since, operates at stage four. It changes how the model treats a page it has already been handed. Stages one to three decide whether it is handed the page at all.
The visibility vector in plain terms
The survey argues that "visibility" should not be a single score. It proposes seven separate things you might be measuring, and it labels them as a vector because they move independently. Stripped of notation:
- Discoverability: how often the engine retrieves you.
- Context exposure: how much room you get once retrieved, and how high up.
- Citation: whether you are named or linked.
- Prominence: where in the answer the citation sits, and how often.
- Absorption: whether your content actually shaped the answer, cited or not.
- Fidelity: whether what the engine attributes to you is true to what you said.
- Behaviour: whether anyone clicks, enquires or buys.
Citation and absorption differ. An engine can cite you and ignore what you said, or use your argument and credit someone else. The survey notes that high citation rates can coexist with low factual fidelity. A citation count on a dashboard tells you one component of seven.
A tactic can also push one component up and another down. The survey points to the SAGEO Arena work, which found that rewriting content to win citations can impair retrieval and reranking upstream, for a net loss of visibility. You can make a page more quotable and less findable in the same edit.
Which tactics the evidence does not support
Keyword stuffing is the clearest failure, and has been since the first paper. It scored 17.7 against a baseline of 19.3 in GEO-bench. The survey's broader finding is that generic heuristics of this kind, the sort that get sold as a checklist regardless of industry, "generally fail or perform inconsistently" when moved to a new engine or domain.
The "40% uplift" as a general claim is rated rejected in the survey's own table of claims. The number was real inside a fixed-context test on GPT-3.5 in 2023. It says nothing about whether a Northampton accountant's site, or any accountancy practice chasing the same visibility, gets found by Perplexity in 2026.
Citation-oriented rewrites are flagged as a risk, for the reason above. Treating a citation as proof of credibility is unsupported too: the paper is blunt that citation metrics are poor proxies for it.
On the "no proven tactic" conclusion, the abstract says:
"No reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior." Olivier Martinez, A Critical Survey of Generative Engine Optimization
Which tactics hold up, with conditions
Two things are reproducible across the studies: topical relevance between the query and the document, and position in the context once retrieved. That sounds like a restatement of search engine optimisation, and to a large degree it is. Be the most relevant page and get read first.
The survey rates "extractable evidence", meaning statistics, definitions and quotations that a model can lift cleanly, as having moderate to strong support. The condition is that the effect was measured after retrieval, and that it varied by engine and by subject area. In Aggarwal's own words, the efficacy of these strategies varies across domains.
Structural changes (headings, lists, question-and-answer blocks) show gains in some studies and nothing in others. The survey's reading is that the effect "depends heavily on the specific domain and engine".
Passage extractability, the idea that each section should stand alone so a chunk can be quoted, sits in the same conditional bucket. Google has since said there is no requirement to break your content into tiny pieces.
Third-party mentions are the interesting one. One observational study in the corpus found earned media overrepresented among cited sources, and another found ChatGPT could recognise 99.4% of named startups but surfaced them in only 3.32% of open discovery queries. Being known differs from being recommended. The survey stops short of saying coverage causes recommendation, because nobody has run the controlled test.
Tactics against evidence strength
The survey grades evidence on five levels, from randomised field trials with real traffic (A) down to fixed-context tests with an LLM as judge (E). Most GEO evidence sits at D and E. This is how the common tactics line up, in our reading of the paper.
| Tactic | What the evidence says | Strength | Condition |
|---|---|---|---|
| Keyword stuffing | Below baseline in GEO-bench; does not transfer from SEO | Rejected | None |
| Generic GEO checklists | Fail or perform inconsistently across engines and domains | Weak | None |
| "40% uplift" as a promise | Fixed-context result on GPT-3.5; rejected as a general claim | Rejected | None |
| Topical relevance to the query | Most reproducible driver in the corpus | Strong | Applies at every stage |
| Position once in context | Reproducible; later sources lose share | Strong | Only after retrieval |
| Statistics, quotations, definitions | Moderate to strong, post-retrieval | Moderate | Varies by engine and domain |
| Headings, lists, Q&A structure | Mixed; limited validation | Weak to moderate | Domain and engine specific |
| Citation-focused rewrites | Can lift citation and cut retrieval | Risk | Measure both before shipping |
| Third-party mentions | Overrepresented in citations; causation untested | Suggestive | Observational only |
| Schema markup | Not isolated in any reviewed study | Unknown | Google: no special schema needed |
| llms.txt | Not studied; Google Search does not use it | None | Harmless, not helpful |
Schema and llms.txt get a row each because clients ask about both. Google's own guide, published in May 2026 and last updated in July, says there is no special structured data for AI features and that Google Search does not read llms.txt files. Schema still earns rich results in classic search, which is reason enough to keep it accurate.
What Google said in May 2026
Google published its first dedicated guide to optimising for generative AI in Search in May 2026, a year after a blog post from John Mueller covered the same ground. The guide's position is that AEO and GEO work "is still SEO", that you do not need new machine-readable files, markup or Markdown, and that the usual SEO practices continue to apply.
It also confirms query fan-out: AI Mode and AI Overviews issue multiple related searches across subtopics to assemble an answer. That is why a page can be cited for a sub-question it never targeted. Google warns that spinning up a page per fan-out variation can breach its scaled content abuse policy.
Two further points fit the survey. Generative features can reflect what is said about a brand across blogs, videos and forums, though chasing inauthentic mentions "isn't as helpful as it might seem". And AI feature performance now has its own Generative AI performance report in Search Console.
The guide does not say which of its recommendations moves retrieval. It is advice from the engine's owner, useful and unverified.
Engines disagree with each other, and with themselves
If you remember one number from the survey, make it this: in one audit, 53% of the domains AI Overviews drew on were outside Google's organic top ten, and 27% were outside the top 100. Ranking well in classic search helps, and it is a different thing.
Engines also overlap poorly with each other. A study of Bing Chat and Perplexity found 26% domain overlap across 355 unique domains. A tactic that works for ChatGPT has no automatic claim on Perplexity.
Then there is run-to-run variance. One 45-day audit across four engines found that the set of sources cited on any given day overlapped only 34% to 42% with the day before. Even at temperature zero, repeated runs changed 9% to 28% of outcomes. A single screenshot of ChatGPT recommending a competitor is an anecdote.
This is why the survey's measurement protocol matters more than any tactic in it. It recommends seven to eight repetitions per prompt as a starting point, three to five paraphrases per information need, and several named engines recorded with version and date. The times the engine did not search at all count as outcomes.
Anyone selling you GEO tracking should be able to say how many runs sit behind each number.
Reddit and the sources that keep getting cited
The survey does not mention Reddit. Commercial data does, repeatedly. Peec's March 2026 analysis of 30 million cited sources in the US put Reddit in the top five domains on all five engines it tracked, first on Gemini and Perplexity and second on Google AI Mode, AI Overviews and ChatGPT.
YouTube was top five everywhere too. Wikipedia led on ChatGPT and was absent from Google's top five.
Read with the survey in mind, that is a retrieval finding. Those domains clear stages one to three constantly. For a UK SME it means an honest, specific answer in a relevant subreddit or forum thread may be retrieved where your own blog post is not.
That is a reason to show up in those places as yourself, and a reason to be wary of any agency offering to seed them.
The click problem
Even when everything works, the click may not come. Pew Research tracked 900 US adults in 2025 and found users clicked a result on 8% of searches that showed an AI Overview, against 15% without one. Only 1% clicked a link inside the overview itself.
The survey does not use the phrase zero-click, but its final section raises the same risk, that summaries substitute for visits. Only a handful of the 45 studies observed behaviour at all.
The one controlled traffic study in the corpus found ChatGPT referrals rose 1.82 times after optimisation, with a placebo test that did not reach significance. So the honest answer to "does GEO bring customers" is: possibly, measured once, not yet replicated.
What this means for a UK SME budget
Spend in the order the pipeline runs, and refuse to pay for stage-four polish on pages that never reach stage three. For a business with a few thousand pounds a month split across SEO, Google Ads and GEO, that means:
- Fix retrieval first. Crawlability, indexing, page speed and genuinely relevant pages are the conditions for everything else, and they are ordinary technical and on-page SEO. The survey and Google agree on this, which is rare.
- Put real evidence in the pages you already rank for. Named statistics, dated quotes from people with titles, clear definitions. This is the one post-retrieval tactic with moderate to strong support, and it costs editorial time, not tooling.
- Measure before you optimise. Pick 20 to 30 prompts your customers would actually type, run each seven or eight times across two or three engines, and record who gets cited. Do it again in a month. If a supplier cannot show you variance, they cannot show you change.
- Earn mentions where buyers already talk. Trade press, directories, forums, Reddit, YouTube. The evidence is observational, but it points the same way as the retrieval data.
- Decline anything sold as a universal checklist, a promised percentage uplift, or a schema and llms.txt package. None of those has evidence behind it, and the last one bundles two things Google says it does not use.
At Growthcart we have moved our own GEO work to that order, and we report citation share per engine with the number of runs behind each figure instead of a single visibility score.
The survey's author is careful to say his conclusion "does not invalidate GEO". It defines the work still to be done. The field is three years old and the first proper field trial has barely been run.
In the meantime, the decision rule we would offer a UK business owner is this: if a tactic would also make the page better for a person reading it, do it. If it only makes sense for a machine, ask for the evidence, and expect there to be none yet.
If you want a GEO audit that starts with retrieval and reports per-engine, repeated-run data, our generative engine optimisation service is built on exactly that, or get in touch and we will talk you through what we would measure first.