Turns Out · · 16 min read
Generative engine optimization (GEO): what it is and the 12 things that actually get you cited
How ChatGPT, Perplexity and Google AI pick sources, 12 on-page and off-page signals that earn AI citations, and how to measure GEO results month by month.
Somebody on your team has already said the word. A vendor deck called it a "new discipline", an agency quoted a monthly fee for it, and the acronym has quietly replaced "SEO" in a few job titles. Generative engine optimization, GEO, is the practice of getting your pages used and cited when ChatGPT, Perplexity, Claude or Google's AI Overviews answer a question your customer asked. The question we get from clients is simpler than the pitch decks: what, specifically, is different from the SEO we already pay for?
Less than the decks say, and more than Google admits. We spent this autumn reading the engine documentation, the two peer-reviewed papers we found on the subject and the largest citation studies we could find, then checking the findings against the sites we audit. This is the generative engine optimization guide we wish we could have handed those clients in January. It covers the mechanism, 12 signals you can test for on your own pages, the short list of things not worth doing, and a way to measure results that does not involve trusting a dashboard you cannot inspect.
How generative engine optimization differs from AEO and SEO
The three terms get used interchangeably, which is how you end up paying for the same work twice. One line each:
- SEO — earn a position in a ranked list of links. The reader picks. Success is a click.
- AEO (answer engine optimization) — be the one source an engine lifts a direct answer from: a featured snippet, a voice reply, the first sentence of an AI Overview. Success is being quoted, with or without a click.
- GEO (generative engine optimization) — be one of the sources an engine reads, synthesizes and cites when it writes a multi-paragraph answer across ChatGPT, Perplexity, Claude, Gemini and Google AI Mode. Success is a citation, and the brand mention next to it.
The overlap is large, and the differences are not where the vendors put them. All three need a crawlable, indexed, fast page with a clear answer on it. The differences are in what the engine does after it fetches the page, and that is where the research gets interesting.
How generative engines actually choose sources
There are two retrieval regimes, and treating them as one is the most common GEO mistake we see.
The first is Google's. AI Overviews and AI Mode draw from the ordinary Google index. The AI features documentation says a page must be indexed and eligible to appear with a snippet, and then states plainly that there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" — no special file, no special markup. What is different is the query fan-out: the system issues "multiple related searches across subtopics and data sources" and assembles supporting links from all of them, so a page can be cited for a sub-question the user never typed. Google's newer guide to optimizing for generative AI features adds one administrative condition: the site must be included in Search generative AI features in Search Console, a setting that defaults to Include and has been available to every property since 31 August 2026. The same guide lists what you can ignore for Google Search: new machine-readable files, chunking, writing in a special way for AI, and special schema.
The second regime is live retrieval by assistants. ChatGPT search, Perplexity and Claude run their own search crawlers, fetch candidate pages at answer time, and read them as text. OpenAI's bot documentation is explicit that OAI-SearchBot is the crawler that surfaces sites in ChatGPT search, that GPTBot is the training crawler, and that "each setting is independent of the others". Perplexity documents the same split between PerplexityBot for its index and Perplexity-User for on-demand fetches, and Anthropic does too with ClaudeBot, Claude-SearchBot and Claude-User.
Two findings tell you how the regimes behave. Ahrefs ran 15,000 long-tail prompts through Google, Bing and four assistants in mid-2025 and found that only about 12% of the URLs the assistants cited ranked in Google's top 10 for the same prompt. That average is pulled up by Perplexity at 28.6%; ChatGPT's citations were at 8% or lower. Ahrefs' own explanation is fan-out, not indifference to rankings: an assistant searches several variations of the prompt and may never search the prompt itself, so a page is often cited because it ranks, for a query nobody typed. The assistants also lean on search engines more than their crawler names suggest: Copilot draws on Bing, ChatGPT has been observed using Bing and Google results, and Perplexity keeps its own index.
Google's own AI Overviews are moving the same way. Ahrefs measured 76% of their citations coming from the top 10 in July 2025 and 38% in March 2026, across 863,000 keyword results. The two runs are not like-for-like (the first counted only the three most visible citations in each Overview), but the direction matches the assistants: ranking for the exact query explains less than it did, and ranking for the questions around it explains more.
The second finding is older, and it is where the term comes from. The GEO paper by Aggarwal and colleagues (KDD 2024) coined the term and tested nine content edits across a 10,000-query benchmark. Adding quotations and adding statistics raised a page's share of the generated answer the most, with fluency edits and citing sources close behind; adding keywords scored below the untouched baseline (17.7 against 19.3 on the paper's position-adjusted word count). The headline number, visibility boosted by "up to 40%", is the one the vendors quote. Read it with its limits: the engine was a simulation (the top five Google results for each query, answered by GPT-3.5 and scored by share of the answer's words), with a smaller check on Perplexity that showed gains of up to 37%. The keyword-stuffing result is the one that should change how you write.
“Compared to baselines, simple methods like Keyword Stuffing traditionally used in SEO don’t perform well.”
The paper has since been re-tested. C-SEO Bench (NeurIPS 2025) ran this family of content edits across more domains and found most of them "largely ineffective" and frequently harmful to a document's place in the answer, while improving where the source ranks in what the model is given was "significantly more effective". The gains also shrank as more competitors adopted the same edits. Our reading: sourced numbers and quotations make a page worth citing, but no rewrite substitutes for being retrieved in the first place.
One more number, because it reorders the usual priority list. Ahrefs correlated AI Overview mentions with site metrics across 75,000 brands: brand mentions on the web correlated at 0.664, branded anchors at 0.527, branded search volume at 0.392, and raw backlink count at 0.218. Correlation is not causation, as the authors say, but if you have a fixed budget for "authority", the engines are telling you where they look.
The 12 signals that get you cited
We grouped them by who owns the fix. Each one is written so you can test it on a page this afternoon. They do not all rest on the same evidence. Three have data behind them: statistics and quotations, on-page sources and third-party mentions, from one simulation and one correlation study. Two are mechanical: an engine that cannot fetch or read the text cannot cite it. Two come from Google's documentation. The other five (answer-first paragraphs, entity clarity, Q&A formatting, llms.txt and brand-name consistency) are reasoning from how retrieval works and from what we see in audits; treat them as hypotheses to check with the measurement method below.
On the page: the engine has to find the answer
- Answer-first paragraphs. The first sentence under every heading states the answer; the rest supports it. Test: read only the first sentence of each section and see whether a stranger could answer the heading. Fix: move the conclusion up and the preamble down. This is the edit we would make first, because retrieval works on passages, not whole pages.
- Entity clarity. The engine should never have to guess what you are, where you are and what you sell. Test: paste your home page text into any assistant and ask it to describe your company in two sentences; if it hedges, so will its answers to your customers. Fix: say the plain thing once, early, in the same words on every page.
- Q&A formatting where questions exist. Headings phrased the way people ask, with a two-to-four-sentence answer under each. Test: do your h2s contain the words "how", "what", "why", "vs" or a number? Fix: rewrite the headings; keep the body. Do not bolt a generic FAQ onto every page — the engines fan out to real questions, not padding.
- Statistics and quotations with the number left in. This is the GEO paper's top result and the one most marketing copy fails. The later benchmark found smaller effects, so add numbers because they make the page worth quoting, not for a guaranteed lift. Test: count the specific numbers and attributed quotes on your best page. Fix: add the ones you can source; strike the adjectives they replace.
- On-page sources. Link to where your numbers came from. Citing sources was among the top-performing edits in the paper, and it is the cheapest to add. Test: does every statistic have a link? Fix: a Sources list at the end, like the one on this page.
- Visible dates and a real author. Google's publication date guidance asks for a prominent, user-visible date and matching
datePublished/dateModifiedin structured data. An anonymous "Admin" byline is a weak signal for anything that claims expertise. Test: find the date and author on your page in under three seconds. Fix: show both, link the author to a page with credentials.
In the markup and the stack: the engine has to read it
- Structured data that matches the visible text. Google lists this among the SEO basics that carry over to its AI features, and says no special schema is needed. For the assistants we know of no study either way; JSON-LD is in the HTML they fetch, so at minimum it is a second, unambiguous statement of who wrote what and when. Test: Schema.org validator, then a diff against the rendered copy. Fix:
OrganizationwithsameAs,ArticleorBlogPostingwith author and dates,ServiceandProductwhere you sell them. We wrote up the trap of mistaking the Rich Results Test for a validator in Google can't see your FAQ schema. - Crawler access that distinguishes training from search. Blocking
GPTBotkeeps your content out of training sets and has no effect on ChatGPT search; blockingOAI-SearchBotremoves you from ChatGPT answers. Test:curl -A OAI-SearchBot https://yourdomain.com/robots.txtand read it, then check your CDN's bot rules, which override robots.txt silently. Fix, if you want to be cited but not trained on:
User-agent: GPTBot
Disallow: /
User-agent: ClaudeBot
Disallow: /
User-agent: OAI-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Claude-SearchBot
Allow: /- Server-side rendering of the content that matters. Vercel and MERJ analyzed AI crawler traffic across their network and found, in December 2024, that none of the major AI crawlers — OpenAI's, Anthropic's, Perplexity's — render JavaScript; ChatGPT's and Claude's fetch the script files and never execute them. The exceptions in that study were Google's Gemini, which rides on Googlebot, and Apple's Applebot. It is the last network-wide test published, so re-run the check below on your own pages. Since October 2026 at least one published log test shows OAI-SearchBot running a page's script from OpenAI's IP ranges; OpenAI has confirmed nothing, the evidence covers a handful of sites, and nobody has shown it for the other crawlers. If your pricing, FAQ or product specs are injected client-side, the assistants are citing your competitor's static page instead. Test:
curl -s https://yourdomain.com/page | grep "a sentence from the page". Fix: SSR, static generation or at minimum a server-rendered fallback for the content blocks that carry the answers. The same fix moves the LCP numbers we covered in Core Web Vitals in 2026. - An llms.txt that is a map, not a brochure. No engine has publicly committed to reading it, which we were blunt about in llms.txt explained. We still ship one, for the reason you are reading this page: a curated list of your best URLs with one-line descriptions costs 30 minutes, cannot hurt, and is exactly the index an agent wants when it lands on your domain. Test: does yours list stale URLs or marketing copy? Fix: facts, current links, nothing else.
Off the page: the engine has to trust you
- Third-party mentions in the places the engine already reads. The 0.664 correlation above was with brand mentions, not links; backlink count managed 0.218. Test: search your brand name in quotes alongside your category and see who, other than you, says what you do. Fix: directories that matter in your industry, partner pages, podcasts with transcripts, guest answers on forums the engines index. Earn them: Google's guide says seeking inauthentic mentions "isn't as helpful as it might seem".
- Brand-name consistency. One spelling, one legal form, one description, everywhere — your site, your social profiles, your partner listings, your
Organizationschema. An engine has only the text to go on; "Acme", "ACME Inc." and "Acme Group" risk being read as three weaker entities instead of one. Test: search each variant. Fix: pick one and go correct the rest, starting with the pages you do not control.
What is not worth doing
The honest list, from what we have seen go wrong on client sites this year:
- Rewriting for keyword density. The paper measured it; it scored below doing nothing. The engines are summarizing meaning, and the stuffing reads as noise.
- Publishing AI-generated "GEO content" at volume. Google's AI features draw on the same index as the ten blue links, so the scaled-content rule in its spam policies covers them too, and a generated answer has no reason to cite a page that says nothing its own model could not have written. Fifty thin pages give an engine fifty chances to cite someone else.
- Blocking every bot that starts with an AI company's name. A well-meaning "block AI" rule in the CDN removes a site from ChatGPT search and Perplexity just as quietly, and Cloudflare has blocked AI crawlers by default for new sites since July 2025, so many owners never made the decision at all. Block the training crawlers if that is your policy; leave the search ones alone.
- Buying a GEO tool before you have a prompt set. The tracking tools are useful once you know which 40 questions matter; before that they produce impressive charts of noise.
- Treating one month of citation data as a trend. Assistant answers vary between runs of the same prompt. A single check is an anecdote.
The caveat that favors the other side: Google is right that for its own AI features there is nothing to do that is not already SEO. Its guide puts llms.txt files, chunking, AI-specific rewriting and special schema on the list of things you can ignore for Search; where our 12 include them, it is for the assistants and for agents, not for Google. If your traffic is overwhelmingly Google and your pages already rank, half of this list is maintenance you are already doing. The 12% overlap number is what should make you care about the other half.
How to measure GEO results
A method you can run without any paid tooling, and then automate when the manual version has taught you what matters:
- Build a prompt set. 30 to 50 questions your buyers actually ask, split across awareness ("what is X"), comparison ("X vs Y", "best X for Z") and purchase intent ("X agency in your city"). Pull them from sales calls, support tickets and the questions in your search console data, not from a keyword tool.
- Pick the engines you care about. ChatGPT with search on, Perplexity, Claude, Gemini, and Google AI Mode or AI Overviews for the prompts that trigger them. Five is plenty.
- Score every answer three ways. Mentioned (your brand appears in the text), cited (your URL is in the sources), recommended (you are named as the answer or in a shortlist). Record which URL was cited — it is often not the page you expected, and that tells you which pages the engines find quotable.
- Record the competitor set in the same pass. Share of voice is the only number that stays meaningful as the engines change.
- Add Google's first-party data. Search Console's Generative AI performance report covers AI Overviews and AI Mode for your own property, so the Google column of the sheet does not have to be collected by hand.
- Repeat monthly, same prompts, same engines, two runs each. Two runs smooth the variance enough to see direction without pretending to precision.
The sheet ends up as prompt × engine × month with three flags per cell. After three months you will know which of the 12 signals moved a citation and which did not, for your site, which is the only version of that answer that is true.
Questions readers ask
Is generative engine optimization just SEO with a new name?
For Google's AI features, almost: the eligibility rules are the ordinary ones plus a Search Console setting that is on by default, and Google says no special optimization is needed. For ChatGPT, Perplexity and Claude it is not: they run separate crawlers, fetch pages live, mostly do not execute JavaScript, and cite pages that mostly do not rank in Google's top 10 for the prompt as typed. The on-page work overlaps heavily; the technical and off-page work does not entirely.
Does llms.txt improve AI citations?
There is no public evidence that any major engine reads it today, and we say so in our own write-up of the file. It is cheap, harmless and useful to agents that do fetch it, so we ship one and expect nothing from it. Spend the time on server-side rendering first.
Should I block AI crawlers?
Decide separately for training and for search. Blocking GPTBot, ClaudeBot and similar keeps your content out of model training and does not affect whether you are cited. Blocking OAI-SearchBot, PerplexityBot or Claude-SearchBot removes you from those assistants' answers. Check your CDN's bot settings as well as robots.txt; they are the usual culprit.
How long until GEO work shows results?
Google's AI features follow its ordinary crawl and index cadence, so changes can appear within days of a recrawl. Assistants that fetch pages at answer time can pick up a change the next time they fetch the page; we know of no published measurement of how fast that is. Earning citations for prompts where you were absent depends on the off-page signals, and those move on the timescale of mentions, which is months.
If you want a baseline before you touch anything, our free AI visibility audit runs a prompt set like the one above against the major engines. It tells you which of these signals your site is currently failing, including the ones a tool cannot fix for you.