Generative engine optimization: what AI search actually rewards

An AI answer is assembled in four steps, and each one is a point where a page is kept or dropped.

• Fan-out. The question a person types is rewritten into several narrower searches. "Which recorder should I buy for a six-camera shop?" becomes separate searches for channel counts, storage sizing and small-business installation. A page that ranks for the head term but for none of the sub-questions is never fetched, and no amount of on-page work fixes that.
• Retrieval. Each sub-query runs against a conventional web index and the top results are fetched. Google states plainly that AI Overviews and AI Mode use the same Googlebot crawl and the same Search systems as ordinary results, which is why an unindexed page is invisible to them.
• Passage scoring. The fetched pages are split into chunks and each chunk is scored against the sub-query rather than against the original question. This is the step most GEO advice skips, and it is the one that decides which paragraph is quoted.
• Synthesis. The model writes an answer from the passages that survived and attaches citations. It can use a passage without citing it, and it can cite a page it barely used, which is why citation counts and influence are not the same measurement.

We build the same shape of pipeline for our own retrieval, so this is not inference from the outside: a vector store, an embedding model and a re-ranker, scoring chunks against a query. The behaviour is consistent and unglamorous. A section with its own heading and a complete answer inside it scores. A paragraph whose meaning depends on the two paragraphs above it scores badly, because the two paragraphs above it are in a different chunk.

The practical reading: you are not optimising a page for a question. You are optimising a passage for a sub-query you will never see.

The three disciplines overlap heavily, and the boundaries matter less than the mechanisms. It helps to name what each one is optimising for.

• SEO earns a position in a ranked list. The unit is the page, the prize is the click, and the signals are relevance, links, technical health and reader behaviour.
• AEO earns the direct answer. The unit is a short passage that a featured snippet, a voice assistant or an AI box can lift verbatim. Our answer engine optimization guide covers that in depth.
• GEO earns a citation inside a synthesised answer. The unit is a claim the model can verify and attribute. The prize is a mention, a link, or both.

Everything rests on the SEO layer: an engine cannot cite a page it never indexed. AEO and GEO then share most of their tactics, because a passage that is easy to lift is also easy to cite. The difference is scope. AEO targets one question with one answer. GEO targets the whole answer, which is why first-party data, sourcing and freshness weigh more heavily. LLM SEO names the same territory, and our GEO definition states it in a sentence.

No, and the reason is structural rather than a matter of opinion. Every generative engine in wide use retrieves candidate pages from a conventional web index before it cites anyone. Google's AI Overviews draw on Google's own index. Perplexity and ChatGPT search run live retrieval against indexed pages. A page that is blocked, unindexed or buried under duplicates is invisible to all of them.

What is changing is the click. When the answer sits above the results, fewer people scroll to the links, so informational queries send less traffic than they did. The traffic that survives skews commercial: someone who has read the summary and still clicks wants to buy, compare or configure. That shifts what a content programme should measure, from sessions towards citations and qualified visits.

The sensible framing is that GEO is SEO with a second audience. The first audience is still a crawler and a ranking system. The second is a model with a context window and an instruction to cite its sources. Serving the second without the first is impossible. Serving the first without the second leaves your brand out of the answer people actually read.

Nobody outside the engine vendors sees the selection logic, so this is inferred from published research, engine documentation and repeated measurement. The pattern is consistent enough to act on.

• Self-contained passages. A paragraph that answers one question without depending on the paragraph above it is easy to extract and safe to quote.
• Verifiable specifics. Numbers with units, named sources, dates and named people. Retrieval and re-ranking favour pages that show their working.
• First-party data. Facts that exist nowhere else: your specifications, your support answers, your measured results. If a model wants that fact, you are the only citation available.
• Freshness. Retrieval layers weight recent updates for time-sensitive queries. A visible reviewed date and a genuine refresh cycle both help.
• Structure. Question-shaped headings, a FAQ block, tables for comparisons, and JSON-LD that states author, organisation and date in machine-readable form.
• Crawler access. OAI-SearchBot, PerplexityBot and their peers must be allowed in robots.txt and served real HTML, not a JavaScript shell. Our free AI crawler check tests that in a few seconds.

None of these is exotic. They are the habits of a careful technical writer, applied deliberately and measured afterwards.

"AI search" is not one channel. The engines retrieve differently, publish different crawler names, and give you different amounts of control.

• ChatGPT with web search. OpenAI documents three separate agents: OAI-SearchBot, which "is used to surface websites in search results in ChatGPT's search features"; ChatGPT-User, which fetches a page when a reader asks about it; and GPTBot, which crawls for foundation-model training. Blocking GPTBot does not remove you from ChatGPT's search results; blocking OAI-SearchBot does. ChatGPT cites few sources per answer and leans towards established vendors and documentation. In our own analytics the links it cites arrive carrying a utm_source=openai parameter, so ChatGPT-referred sessions are separable without guessing.
• Perplexity. PerplexityBot surfaces and links websites and, in Perplexity's own words, "is not used to crawl content for AI foundation models"; Perplexity-User fetches on a reader's behalf. Perplexity cites many sources per answer and leans on comparison articles and review sites, which makes it the widest surface and usually the first one a new site appears in.
• Google AI Overviews and AI Mode. No separate crawler and no separate optimisation. Google's documentation says: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and "You don't need to create new machine readable files, AI text files, or markup to appear in these features." The controls you have are the ordinary ones: nosnippet, data-nosnippet and max-snippet. Google-Extended governs whether your content trains Gemini; it does not remove you from AI Overviews, which are served by Search from Googlebot's crawl.

On 31 August 2026 we put the same twenty curated buyer questions to both engines for four sites we publish for. Perplexity cited those sites in 12, 7, 5 and 2 answers out of 20. ChatGPT's web search cited them in 10, 5, 3 and 0. Same pages, same questions, same day. If a tool gives you one number for "AI visibility", ask which engines are in it and in what proportion.

Most GEO measurement is wrong in one of five ways. We have made four of these mistakes ourselves and fixed them, which is the only reason we can be specific.

• Probing without retrieval. Our own visibility probe asked the model a plain question — "list the top suppliers for X" — and read the answer. That answers from parametric memory, so only brands already famous enough to sit in the model's weights can score. It returned 0 citations from 80 probes across four sites, for five consecutive weeks. With the web-search tool enabled, the same queries on the same day named two of those sites first and second, with citation annotations pointing at their own pages. Ask any vendor whether their score uses the assistant's search mode or its bare model. The answer changes the number completely.
• Probing your own brand name. Ask a model to list suppliers for a query that is your brand name and it will name your brand. Our June figures looked healthy for exactly that reason, and the honest number after we curated the query list was zero. Vanity, then honest, then correct — in that order.
• Counting links instead of answers. An engine quoting one publisher three times in one answer is one endorsement, not three. Share of answers is the metric; share of links rewards whoever happens to be quoted repeatedly.
• Recording a failed probe as a zero. "We could not measure" and "nobody was cited" are different claims. Storing an error as a zero manufactures a decline that never happened.
• Measuring once. Answers vary between runs, between signed-in readers and between countries. One check is an anecdote; a fixed basket asked on a fixed schedule is a series.

Cadence is a cost decision, not a preference. One search-backed ChatGPT probe costs roughly 8,800 input tokens plus a billed web search at about two cents, and takes about a minute of wall clock. Running twenty questions per site across four sites every night would be about eighty minutes and roughly $54 a month, to re-measure something that moves in weeks. We run the ChatGPT probe weekly and Perplexity nightly, and on the other days the ChatGPT probe is skipped rather than downgraded to the parametric ask, because two different measurements in one column is what caused our confusion in the first place.

The published version of this method is our AI Citation Index: twenty-four questions, asked every Monday, every cited source recorded, our own domain in the table on the same terms as everyone else's.

GEO advice tends to assume the page is in the index. Often it is not, and every hour spent on quotable passages before that is wasted.

On 25 August 2026 we measured, for four sites the engine publishes to, what share of published articles Google had indexed: 25 of 27, 8 of 13, 14 of 28, and 12 of 25. Same pipeline, same schema, same cadence. On the worst site, eleven of the thirteen misses were reported by Search Console as "URL is unknown to Google" with an empty last-crawl date — never discovered, never fetched. We ruled the usual suspects out one at a time: all were in the sitemap, the sitemap was downloading cleanly with no errors, robots.txt allowed them, they returned 200 to Googlebot, they carried "index, follow" and a self-referencing canonical, and they were reachable through crawlable pagination. Nothing was blocking them. Google simply had not spent the crawl budget.

So when pages do not appear, check discovery before quality. "Unknown to Google" is a crawl-demand problem, and the remedy is a fresh sitemap lastmod, internal links from pages that are crawled often, and IndexNow — not a rewrite.

One trap worth naming, because it cost us months of false confidence: Google's Indexing API accepts URLs it will never index. urlNotifications:publish returns HTTP 200 with no latestUpdate field, and when we queried the metadata endpoint for every URL we had "successfully" submitted across five sites, every one returned 404 NOT_FOUND — owned, accepted, never recorded. A control property we do not own returned 403, so the endpoint does distinguish ownership. Google scopes that API to JobPosting and BroadcastEvent pages, and our pages are neither. A 2xx is not an answer; read the body.

• Check: what percentage of your published pages does Search Console report as indexed?
• If it is below about 80%: discovery and crawl budget are your GEO problem, not passage structure.
• Then, and only then: work on what a retrieval layer does with the pages that are in.

This is the sequence we run for our own tenants, stripped of anything that only works with our tooling.

• Measure before you touch anything. Take the twenty questions your buyers actually ask, put them through ChatGPT, Perplexity and Gemini, and record who gets cited. Our free AI Visibility Checker gives you a first read for a domain.
• Clear the crawl. Confirm the AI crawlers can fetch your pages and that the content is in the HTML, not injected after load.
• Choose questions, not keywords. Search Console queries, People Also Ask boxes and forum threads tell you the phrasing a model will be asked in.
• Write the citable unit. Open each section with a one-sentence definition or answer. One claim per paragraph. A source or a number behind every claim that could be challenged.
• Put your own data in. Datasheets, support tickets, warranty rates, install notes. This is the part a competitor cannot copy.
• Mark it up. Article, FAQPage and Organization schema, a named author with a real biography, a visible reviewed date.
• Refresh and re-measure monthly. Citation share moves slowly and noisily, and a monthly sample of the same queries is the smallest series that shows a trend.

IT Master is an autonomous content engine, and GEO is part of how it produces every article rather than a service bolted on afterwards.

Topics come from real demand: DataForSEO and the customer's own Search Console data. Research pulls competitor pages, forum threads and, where it exists, the customer's own first-party data, so the draft carries facts a rival site cannot reproduce. On the live UK CCTV trade and retail sites, that data is product detail, supplier datasheets and support questions.

A Claude model writes. Models from a different vendor then check the draft: Gemini and GPT fact-check the claims, score novelty against the competing pages, critique E-E-A-T and look for AI tells. A draft can go round up to three times. Across 186 Standard-tier runs, 21% of drafts cleared every check on the first attempt with no revision, and 51% got there within the three revisions allowed; the rest never published.

Passing articles publish with JSON-LD, internal links, a hero image and a FAQ block, and IndexNow pings the search engines. A nightly loop refreshes decaying pages and samples GPT and Perplexity for citations. Details are on the GEO services page and in how it works; you pay per published article, see pricing.

Nobody controls what a model says. That sentence should sit above every GEO proposal, including ours.

A generative engine can paraphrase your page inaccurately, attribute your data to a competitor that quoted you, or produce a perfectly good answer with no citations at all. Its output varies between runs, between signed-in readers and between countries, so any citation figure is a sample, not a rank. Two people asking the same question on the same day can see different sources.

Citation is not traffic. A brand can be named in every answer for a topic and still receive fewer clicks than it did from position three in the old results. Decide what you want the mention to do before you invest in getting it.

GEO also does not rescue thin content. The engines are looking for evidence and expertise, and a page that adds nothing to what they already hold will be passed over however well it is structured. The tactics here make a good page citable. They do not make a weak page good.

Finally, AI answers do not appear on every query, and the engines change their retrieval behaviour without notice. Build the measurement first, so you know when the ground moves.

Primary sources for the claims on this page. Where we quote a vendor, the quote comes from the page listed here, and it was read on 4 September 2026.

• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, "GEO: Generative Engine Optimization", submitted 16 November 2023 — the paper the term comes from, and the source of the "up to 40%" figure: arxiv.org/abs/2311.09735
• Google Search Central, "AI features and your website" — that AI Overviews and AI Mode use the ordinary Googlebot crawl, and that no separate optimisation or AI text file is required: developers.google.com/search/docs/appearance/ai-features
• Google Search Central, "Robots.txt specification" — that a crawler obeys only the most specific matching group and ignores the rest: developers.google.com/search/docs/crawling-indexing/robots/robots_txt
• OpenAI, "Bots" — OAI-SearchBot, ChatGPT-User and GPTBot, and what each is for: developers.openai.com/api/docs/bots
• Perplexity, "Bots" — PerplexityBot and Perplexity-User: docs.perplexity.ai/guides/bots
• Google Search Central, "Indexing API quickstart" — the API is scoped to JobPosting and BroadcastEvent pages: developers.google.com/search/apis/indexing-api/v3/quickstart
• IndexNow protocol documentation: www.indexnow.org/documentation
• Our own measurements: the AI Citation Index, published weekly under CC BY 4.0, and the AI Content Index, read live from the engine's job ledger.