Most of what is sold as a playbook for GEO, or generative engine optimization, fails when retested. C-SEO Bench (NeurIPS 2025) re-tested ten conversational-SEO rewrite methods across two tasks and six domains and found that most of them were largely ineffective and often harmful to citation rank, with gains shrinking as more sites adopted them. What is left is fewer and plainer, and it is the playbook this article sequences: pages engines can retrieve, evidence engines can quote, third-party sources that carry your name, and an entity engines can resolve.
The short answer to the question in the title comes in four moves. Serve your buyer-question pages as raw HTML that AI crawlers can read, load them with sourced statistics and named claims, earn presence in the third-party sources each engine already cites, and describe your brand the same way everywhere it appears. The GEO study (Aggarwal et al., KDD 2024) measured a 30 to 40% citation-visibility lift from added evidence, the page-level tactic family with the strongest published support.
The stakes sit in the shortlist. G2's research 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 a third of buyers said an AI chatbot led them to choose a vendor they hadn't considered. An engine that never cites you is shaping those decisions without you.
How do AI engines decide which brands to cite?
Every engine assembles an answer in two stages, retrieval then synthesis, and a brand can drop at either one. For current commercial questions, ChatGPT search rewrites the request into one or more targeted queries and reads what comes back, and that retrieval leans on conventional search infrastructure: Seer found 87% of SearchGPT citations matched Bing's top results. If nothing of yours is retrievable for the question, the citation is decided among the pages that were.
Synthesis then weighs what was retrieved against what the model already associates with your category. AirOps' March 2026 analysis of 548,534 pages ChatGPT retrieved found only 15% were cited in the final answer, and a 2026 analysis of brand dynamics in LLM recommendation systems found well-known brands were recommended every time only when all products had identical specifications, an edge that disappeared when a competitor had a rating advantage of less than 0.1 stars. The audience is not small: ChatGPT reached 800 million weekly users by spring 2025.
Be precise about what you are chasing. A mention is your name in an answer, a citation is your page used as a source, a recommendation is the engine advising a buyer to choose you, and a shortlist position is surviving an agent's comparison. 6sense found buyers evaluate roughly five vendors and most of the list is set before contact, so the higher rungs carry the revenue. Citations are the rung this playbook builds, and the others stand on it.
Which GEO tactics survive retesting and which do not?
Far fewer survive than the playbooks claim, and the corrective is load-bearing for everything else in this piece. The original GEO benchmark reported that adding statistics, quotations, and source citations lifted citation visibility by roughly 30 to 40%, with its strongest configuration improving on baseline by 41%. The same paper's error-bar table showed keyword stuffing did nothing.
A later benchmark then cut deeper. C-SEO Bench re-tested ten conversational-SEO methods and found most of them largely ineffective and often harmful to citation rank. That result demotes page rewriting from a growth lever to table stakes: worth doing because evidence is what engines quote when they quote you, and unsafe to budget as a guaranteed lift.
Some tactics test below zero. Scarcity and exclusivity framing measurably reduces how often an LLM recommends a product, a result established on three sets of 10 fictitious products so brand familiarity could not explain it. And injections that do move answers, like the social-proof cues worth +10.60 points in one evaluation, belong in your threat model: a gain that exploits a model quirk is a gain the platform will patch.
The honest rule: any GEO tactic without an error bar is a hypothesis, and the four gates in the rest of this piece are the ones with published evidence behind them.
How do you make your pages retrievable to AI crawlers?
Serve substance as raw HTML, because most AI crawlers do not execute JavaScript. Crawler research Vercel ran with MERJ documented that GPTBot and its peers read the un-rendered page, so client-rendered content is an empty shell to them, and Cloudflare's agent-readiness scan of the 200,000 most visited domains found that while 78% had a robots.txt, only 4% declared AI usage preferences and 3.9% served Markdown to agents that asked for it. A site can rank on Google and still be unreadable to the engine your buyer asked.
The crawlers are not interchangeable. OpenAI documents four relevant user agents with different jobs, and a robots.txt rule that blocks the wrong one can take you out of live answers when you only meant to opt out of training. The volume is compounding: Cloudflare tracked GPTBot rising from 2.2% to 7.7% of crawler request share in a year.
The audit is mechanical and takes an hour. Fetch your five most important buyer-question pages with JavaScript disabled and read what survives. Then check robots.txt and your CDN bot rules against the crawlers you want, and confirm your core claims sit in the HTML rather than behind a script or an interaction.
What makes a page worth quoting once an engine reads it?
Extractable evidence: specific claims with numbers and named sources, placed where the answer begins. An engine quoting a page needs a sentence that stands alone and says something checkable, with the attribution attached. Open every section with the direct answer and support it afterward; preamble is what engines skip.
Build the evidence before the adjectives. The lift the GEO benchmark measured came from added statistics and quotations with sources, and the claims have to be ones you are entitled to make: a number you can defend to the source, a comparison with stated criteria. Evidence you cannot defend is a correction waiting to be quoted back.
Structure follows the buyer's question. One page per real question beats one page per keyword cluster, because engines retrieve against the question asked. Keep the claim, its number, and its source inside the same paragraph so extraction preserves attribution.
How do you win the citations you do not control?
Earn presence in the specific sources each engine already trusts, because your own site holds only part of the answer: Profound's citation research found more than 4 in 10 AI citations point to sources brands do not control, including editorial coverage and community threads.
The mix differs by engine, so read real answers rather than assuming. Ahrefs' analysis of 4 million AI Overview URLs maps one engine's source profile, finding 38% of citations rank in the organic top 10, and Semrush's AI Visibility Index, built from 126 million AI search prompts, exists because brands surface differently across platforms. The target list that matters is the one inside your own category's answers.
The method takes one afternoon. Run your buyer questions, open every source the answers cite, and list the domains that recur. Those recurring third parties, ranked by how often they carry your category's answers, are your earned-media queue, and presence in the top two moves more than another homepage edit.
Entity consistency is the quiet half of this gate. A brand that appears under three names, or describes itself differently on every profile, fragments into an ambiguous string that readers and engines can pin on the wrong company. Use one name and one description everywhere a profile describes you. The question volume this feeds keeps growing: in Profound's 7.5 million-conversation en-US sample, the share of ChatGPT conversations classified commercial rose from 13.9% to 19.2% in a year. Profound's "more than doubled" volume figure is an extrapolation from estimated weekly users.
How do you measure citation share without fooling yourself?
With repeated runs and base rates, because single answers are noise. SparkToro and Gumshoe's research found AI engines are highly inconsistent when recommending brands, and Search Engine Land's coverage of the same study reported that AI recommendation lists rarely repeat exactly. One screenshot is a sample, and a sample is where measurement starts rather than ends.
The statistics are unforgiving on this point. A 2026 variance-components study found run-to-run noise large enough to swamp real differences in small samples, and work on quantifying uncertainty in AI visibility reaches the same verdict with confidence intervals. So ask each of five to ten buyer questions in two or three phrasings on every engine, weekly for a few weeks, and record who was mentioned, who was cited, and who was recommended as separate columns.
Resist the single score. The IAB's measurement guidance keeps Presence, Prominence, Portrayal, and Persuasion as four distinct things to track, and collapsing them into one number hides which gate is failing. A citation-share trend per question, per engine, tells you where to work next.
How does Trovance run this playbook as a system?
Trovance operates the loop above every week instead of as a quarterly project. You define the tracked questions your buyers actually ask, and it sweeps them across AI engines, preserving every answer run as a snapshot with its full context: who was mentioned, who was cited, who was recommended, and which sources carried the answer. Share of answer then shows where you stand per question, per engine, over time, which is exactly the base-rate measurement the variance research demands.
The snapshots make the four gates diagnosable. A brand that never appears while its pages resist a raw fetch has a retrieval problem; a brand that appears but loses to a competitor's quoted numbers has an evidence gap; answers built on third-party sources that omit you point at earned coverage; answers describing a different company point at entity resolution. Because runs are compared across cycles, you can see which gate moved after you shipped a fix, and decide the next one from evidence.
Acting on the diagnosis draws on the Brain, which keeps the claims you are entitled to make with the proof behind each one. Recommended actions arrive as a ranked queue of content worth making, backed by answer evidence. Drafts are written in your voice from your author voice guidance and research sources, and a person reviews and approves everything before it publishes; the next analysis cycle then reruns the same questions to verify whether the answer actually changed.
What Trovance will not promise is a citation. No honest system can, because engine outputs are probabilistic, competitors keep publishing, and the test results above show how few tactics move answers on command. What it will do is keep the evidence current, so every decision is steered by this month's answers rather than last quarter's screenshot.
What should you do this week?
Run the gates in order, cheapest first. Fetch your five most important pages with JavaScript disabled and fix what disappears. Pick five to ten real buyer questions and establish a base rate across at least two engines. Add one defensible, sourced statistic to each page that answers a buyer question, then list the third-party domains your engines cite most and start earning the two that recur.
Then be honest about timelines. Retrieval fixes can reach answers once engines re-fetch the page; earned third-party coverage takes longer; and every measurement carries the run-to-run variance documented above. Anyone selling guaranteed citations is selling against the test results.
If you want the loop running without the manual afternoons, start a free Trovance trial and see which of the four gates is holding your citations back.
Run the playbook
The four gates of GEO: the gate model this playbook sequences, in full.
AI crawlers don't run your JavaScript: the retrieval gate, audited step by step.
How to build a competitor comparison page: the evidence-first page most categories are missing.
How to show up in ChatGPT without chasing hacks: the same discipline applied to a single engine.
Read the evidence correctly
The most-quoted study in AI search, quoted correctly: what the GEO benchmark did and did not find.
Where AI citations come from: the source mix behind the 57% figure, industry by industry.
How to measure AI search visibility without one score: the measurement protocol in depth.
Social proof is an attack surface, not a GEO tactic: why bias injections belong in your threat model.
