Most of what circulates as LLM optimization has been retested by people with nothing to sell, and the retesting was unkind. When a research team rebuilt the published conversational-SEO rewrite methods and scored them properly, most proved largely ineffective and often harmful to citation rank. That is the finding to budget against, because it separates the techniques worth an engineering ticket from the ones worth a shrug. What remains sorts into two groups, and neither group is about phrasing.
Techniques with the strongest published case add extractable evidence to a page, meaning statistics, direct quotations, and cited sources. The published KDD version of the GEO study defined that set. The three tiers, and what stands behind each:
| Technique | What the evidence shows | Where it comes from |
|---|---|---|
| Statistics and quotations | 41% best case in source visibility | GEO authors' benchmark engine, not a live one |
| Access work | Pays regardless of which ranking study holds | No ranking claim needed |
| Rewriting, evidence edits included | Mostly ineffective, often harmful to citation rank | C-SEO Bench |
G2 found AI chatbots and review sites are now the two biggest influences on B2B shortlists, 51% of software buyers now start research with an AI chatbot more often than with Google, and 6sense found buyers already know 75% of the vendors on their shortlist by the time it is filled. A quarter spent on techniques that move nothing is expensive.
Which techniques have published evidence behind them?
Two, most strongly: adding statistics, and adding quotations attributed to named sources. Both put evidence a model can lift into the part of the page that answers the question. The GEO preprint reports its evidence methods improving on the baseline by roughly 30 to 40% on its word-count measure of source visibility, and 41% for the best single method, with the largest gains in the evidence-bearing edits and smaller ones for stylistic edits such as fluency. That 41% belongs to the authors' own benchmark, and C-SEO Bench's retest found those same methods rarely moved citation rank.
State the conditions or you will misread the result. The evaluation ran on GEO-bench, a query set the authors assembled, scored by a generative engine they constructed over search results. It is not a measurement of ChatGPT, Perplexity, or AI Overviews, and the authors do not claim it is. The KDD proceedings entry and the project's own citation block are the versions to quote.
Effect size varies by domain and by method. Treat the 41% as a ceiling under one stated set of conditions, not a figure to plan against.
What does the retesting say about rewriting?
It says most of it does not hold up. C-SEO Bench took the published methods and ran them across two tasks and six domains, a wider test than the original work applied, and the significant results nearly vanished. The authors corrected their significance tests for the many comparisons they ran, and most methods failed to clear the bar; many significantly lowered citation rank.
The methods were also scored where every competitor applies them, and the advantage collapses once adoption is universal. A technique that only works while rivals ignore it is a timing bet, not a program.
Some rewriting is worse than neutral. Scarcity and exclusivity framing measurably reduces how often a model recommends a product, tested against three sets of 10 fictitious products so brand priors could not explain it. The persuasion language your conversion team likes is a cost on this surface.
Which techniques are about access rather than persuasion?
Three: making the page fetchable as HTML, making the brand resolvable to one entity, and making the answer liftable from a single passage. They pay regardless of which ranking study you believe, because they decide whether your page is a candidate. None of them ask a model to be persuaded of anything.
Make the page fetchable as HTML
Most AI crawlers do not execute JavaScript. Vercel's crawler research with MERJ documented that they read raw HTML, while Google renders pages through its Web Rendering Service, so a client-rendered page can rank on Google and arrive blank at the engine your buyer asked. Cloudflare scanned the 200,000 most visited domains for agent readiness, and its crawler report puts GPTBot's share of crawler traffic at 2.2% rising to 7.7%. Check your robots rules against the four relevant OpenAI user agents.
Serve crawlers the same HTML you serve people. Diverging by user agent is cloaking under Google's spam policy, and it makes your own diagnosis unreliable.
Make the entity resolvable
A brand that appears under three names and describes itself differently on every page hands a model a string to match, not an entity to resolve. Entity-oriented retrieval has been benchmarked across 443 configurations on a news collection, and it is a caution, not a lever: once entity signals could not peek at the answers, none added more than 0.051 MAP. The case for naming is attribution, not retrieval. One name, one description, one canonical URL per page, and consistent references off-site remove an ambiguity the model otherwise settles alone, which is necessary and not sufficient.
Make the answer easy to lift
Put the answer where retrieval can reach it. Liu et al. showed models use information at the start or end of a long context far better than material buried in the middle, and retrieval operates on passages rather than pages. In one 2026 study over 1,186 multi-hop chemistry questions, the gold evidence passages averaged 188 tokens with a 56.26-token standard deviation. That describes how the dataset was built, not how engines chunk pages: commercial engines do not publish their chunk sizes, and there is no target length to write to.
The advice follows from the pipeline's shape, not that measurement. A self-contained paragraph near the top of a section can be lifted whole; a claim that depends on three paragraphs above it cannot, and what actually gets cited is covered separately.
Retrievability gates everything else. Seer found 87% of SearchGPT citations across a 500-citation sample matched Bing's top results, and ChatGPT search issues one or more targeted queries, so the query that reaches a search index is what your page has to match. AirOps' March 2026 analysis of 548,534 pages ChatGPT retrieved found only 15% were cited, and pages whose titles closely matched the query were cited 20.1% of the time, against 9.3% for pages that barely matched.
How do you structure FAQ, table, and snippet content so ChatGPT and Perplexity cite it?
Write the question in the words a buyer would use, answer it in the first sentence, and put the context underneath. Use a table when the answer is a short set of comparable values, as the tier summary above does, because header cells and rows carry labels a retriever would otherwise infer. Mark it up as an HTML table: a screenshot carries no text to retrieve, and a grid that only looks like one carries no header relationships.
Treat format as extraction craft rather than a lever that earns citations. Google states that its AI features need no special optimizations and no new machine readable files or markup, only a page that is indexed and snippet-eligible. That guidance is about Google's AI features; engines that query a search index first put the block through the path above. Established schema types are not the new markup Google says you can skip: structured data describes the page in terms a search system already parses, and Google publishes the types it supports; mark up only what a reader can see.
No result cited here isolates FAQ-versus-no-FAQ or table-versus-prose citation rates, and the retesting pointed at traditional retrieval-side improvements rather than rewriting.
Where does on-page work stop mattering?
At the point where the answer is assembled from sources you do not own. Profound's citation research found more than 4 in 10 AI citations point to sources brands do not control: comparison articles, editorial coverage, community threads. No on-page technique edits a community thread.
That is not an argument against the access work, which is cheap and should happen anyway. It is an argument against treating on-page technique as the whole program.
How do you know whether a change worked?
You measure a rate across repeated runs. A single recheck the day after you publish reads variance as a result. SparkToro's research found AI engines are highly inconsistent when recommending brands, as Search Engine Land's write-up of the same study also reports. A technique that looks effective in one before-and-after check has not been tested.
Size the sample against the noise. A 2026 variance-components study found run-to-run variation large enough to swamp real differences in small samples, and work on quantifying uncertainty in AI visibility does the same with confidence intervals. We treat several weeks of answers, across phrasings and engines, as a working floor; the study prescribes no number.
Measure the right rung too. The IAB's framework separates presence, prominence, portrayal, and persuasion, and those move independently: being mentioned is not being cited, and being cited is not being recommended. The web search tool returns its output in two main parts, one carrying the citation annotations, which is why citation is the rung you can observe directly.
How does Trovance turn these techniques into decisions?
Trovance turns technique selection into an evidence question. You define the market questions your buyers ask, and it sweeps them across AI engines every week, preserving each answer: which rung your brand reached, and which sources built the answer. That record shows whether your page is among the sources an answer cites, and which pages it cites instead.
The record routes you to the right tier. Share of answer that excludes your pages while a competitor's evidence gets quoted points at access and proof, both of which you control. Answers built from sources you do not own point elsewhere. The Brain keeps your claims with their proof, and the ranked queue is backed by answer evidence, so a recommendation starts from what the answers show is missing rather than from a request for another page.
Drafts are written in your voice from your voice guidance and research sources, and a person edits and approves everything before it publishes. After an asset goes live, the next analysis cycle reruns the same questions and compares the new answer snapshots against the earlier ones, separating a real change from variance.
What Trovance will not promise is a citation, a ranking, or a recommendation. The engines are probabilistic, the sources shift under you, and the retesting above is the reason nobody should sell a guaranteed outcome from an on-page edit. There is no single visibility score here either; the rungs are reported separately because they fail for different reasons, and we are building toward tighter attribution between an approved asset and the answers that follow it.
What should you do this week?
Sequence by tier. First, fetch your five most important pages with JavaScript disabled; if what comes back is an empty shell, nothing downstream matters. Second, check your robots rules against the named AI user agents and confirm crawlers get the same page people get. Third, take the two or three pages that answer real buyer questions and put a specific number, a named source, and a direct quotation into the first paragraph under each heading.
Then measure before doing more. Run ten buyer questions across at least two engines, in two or three phrasings each, every week, until you can see each one's spread; that's our working floor rather than a published standard. Record mention, citation, and recommendation separately, and read the sources the answers cite. If most are third-party, your next quarter belongs to earned coverage.
What to skip: rewriting tone for authority, padding keyword density, and any tool that scores your prose against an undisclosed model of what engines want. The published retesting does not support it. If you would rather run the measurement continuously than as a one-off afternoon, start a free Trovance trial and see which tier your gap is actually in.
The techniques and the evidence
The most-quoted study in AI search, quoted correctly - what the GEO benchmark tested.
What gets cited by AI - the passage-level view of evidence.
How to show up in ChatGPT without chasing hacks - tactics that circulate without support.
The four gates a page clears before it is cited - the sequence technique clears first.
Access, entity, and measurement
AI crawlers don't run your JavaScript - the fetchability tier in detail.
A named entity is not a retrieval signal - why entity consistency is necessary and not sufficient.
There is no such thing as an AI visibility score - why the rungs get measured separately.
How to track AI citations - running the repeated measurement.
