A marketing manager publishes a 2,500-word guide on her company’s blog. It ranks on page one of Google for its target keyword. Organic traffic climbs. The classical SEO playbook is working.
The same guide is invisible in ChatGPT Search. Absent from Perplexity’s answers. Missing from Google AI Overviews.
The content is strong. The information is accurate. The page is well-linked. But ranking on a search results page and getting cited inside a generated answer are two different jobs. Figuring out how to rank in ChatGPT, Perplexity, or Google AI Overviews requires work that classical SEO never did, because the unit being ranked shifted from the page to the passage.
This post is the step-by-step process for moving a page from “invisible to AI engines” to “cited as a source in generated answers.” It covers auditing where a site stands today, restructuring content so engines can extract it, and verifying results across the three engines that matter most in 2026.
What You’ll Learn
This guide is for content managers, founders, and SEO practitioners who already publish content and want that content to appear as a cited source when AI engines answer questions in their category.
Readers will finish with six concrete steps. First, how to audit current citation share on target queries. Second, how to pick the queries worth pursuing. Third, how to restructure content so the engine’s passage ranker can find and extract a direct answer. Fourth, how to add the evidence layer that lifts citation probability. Fifth, how to handle the technical prerequisites that determine whether an engine can fetch the page at all. Sixth, how to verify results and know whether the work is moving.
The upstream complete guide to generative engine optimization covers the full GEO framework: what the discipline is, where the term comes from, and the three-layer stack. This post assumes the framework and delivers the hands-on playbook.
How AI Engines Choose What to Cite
The citation pipeline is the same across engines. Four steps, in order.
Retrieve. The engine pulls candidate sources from a web index or a search partner. ChatGPT Search retrieves through Bing. Perplexity combines its own index with Google results and targeted browsing. Google AI Overviews uses Google’s own index. If a page is not in the candidate set, it cannot be cited. This is where classical crawlability still matters.
Rank. Each candidate is scored on relevance, authority, freshness, and a passage-level signal. The engine does not rank the page as a whole. It ranks individual passages within the page against the query. This is the mechanical shift that separates GEO from SEO: the unit of ranking is the passage, not the page.
Synthesize. The model writes the answer by weaving phrases and structure from the top-ranked passages. The synthesis step is more extractive than most writers expect. Engines pull near-verbatim phrases from the passages that ranked highest.
Cite. Sources whose passages contributed to the synthesis appear as citations. Inline in Perplexity and ChatGPT Search. Sometimes at the end of the answer in Google AI Overviews. Not at all in some conversational responses from Claude or Copilot.
The full retrieval-to-synthesis pipeline, including rendering decisions and context-window constraints, sits in the companion post on how AI agents actually fetch and read a webpage.
The operational takeaway: a page that is fetchable, contains well-structured passages, and includes extractable evidence has the highest probability of being cited. The six steps below address each of those conditions in order.
Step 1: Audit Your Current Citation Share
Before optimizing anything, measure where the site stands today.
Pick five to ten questions a target customer would actually ask an AI engine. Not keyword phrases. Full natural-language questions, the way a person types into a chat window.
Examples for a CRM company: “which crm is best for a small agency that needs email sequences and a pipeline”, “how do I choose a crm if I have fewer than ten salespeople”, “what is the best crm for startups in 2026”.
Prompt each question into three engines: ChatGPT Search, Perplexity, and Google AI Overviews. For each response, log three things:
- Whether the site appeared as a cited source
- Which competitor sites were cited instead
- Which passage from the cited source the engine appeared to extract
Run the same five queries once a week for four weeks. The baseline is not a single snapshot. It is the pattern across multiple runs, because AI answers are probabilistic. The same query can return different citations on different days.
For teams with budget, automated citation-tracking tools (Profound, OtterlyAI, Peec AI, AI Monitor, Authoritas, Rank Prompt) run this process at scale. For a team starting out, the manual version is enough and more interpretable.
The audit answers the first question: is the site being cited at all, and if not, which competitors are winning the citations the site should own?
Step 2: Pick the Queries You Want to Win
Not every query is worth pursuing for AI citation. The queries that belong on a GEO target list share two properties.
First, the searcher expects a synthesized answer rather than a specific destination. “Login to Salesforce” is a navigational query that belongs to classical search. “Which CRM is best for a five-person agency” is a synthesis query that AI engines answer directly.
Second, the site has genuine authority to answer the question. An engine choosing citations for “best CRM for small agencies” will favor sites that have published substantive, well-sourced content about CRMs. It will not cite a site that mentions CRMs in passing on its about page.
The starting list comes from the audit in Step 1. Expand it by asking: what are the ten questions a potential customer asks on the way to a purchase decision? What are the five questions they ask after the decision but before implementation?
This is question research, not keyword research. The output is a list of full sentences, not a list of short-tail phrases. The dimension-by-dimension GEO vs SEO comparison walks through why the query model shifted and how it changes the content brief a writer works from.
Step 3: Restructure Your Content for Passage Ranking
This is the step where most of the citation-share gains come from.
AI engines rank passages, not pages. A 3,000-word article that buries its direct answer in paragraph twelve is a weak citation candidate, even if the answer is accurate and the article is well-written. The passage ranker scores each paragraph against the query, and the winning passage is almost always near the top of a cleanly structured page.
Four structural changes move the needle.
Answer the question in the first paragraph. The opening of any page that should earn citations needs to contain the direct answer to the question the page is about. Not the history. Not the context. Not the “why this matters.” The answer.
A page targeting “how to choose a CRM for a small agency” should open with something like: “A small agency choosing a CRM should prioritize three features: a simple pipeline view, built-in email sequences, and per-seat pricing under $30 per month.”
The framing and context follow. The answer comes first.
Use descriptive headings, not clever ones. An H2 that reads “The CRM Landscape” tells the passage ranker nothing about what the section answers. An H2 that reads “How to Choose a CRM for a Small Agency” tells it exactly what to match against the query. Engines use headings as passage boundary signals. Descriptive headings help the ranker match the right section to the right query.
One idea per paragraph. A paragraph that covers three topics is three weak passages instead of three strong ones. Split it. Each paragraph should be a self-contained unit that makes sense if the engine extracts it without its neighbors.
Use lists for enumerations. When a paragraph contains a sequence of items (“the three features to look for are X, Y, and Z”), render it as an HTML list. Engines extract lists cleanly. They struggle with comma-separated enumerations buried in the middle of a dense paragraph.
Step 4: Add Evidence Engines Want to Extract
Passage structure gets a page into the candidate set. Evidence gets it cited.
The Princeton and Georgia Tech paper that introduced the term “generative engine optimization” measured the effect of several content interventions on visibility in AI-generated answers. Two findings have held up consistently as the industry applied them.
Cited statistics lift citation probability. A passage that includes a specific number with a named source (“agencies with fewer than ten users pay a median of $25 per seat per month, according to G2’s 2026 SaaS pricing report”) is more likely to be extracted than the same claim without the number or without the source.
The mechanism is straightforward. The synthesis step is extractive. When an engine needs a factual claim for its answer, it prefers a passage that already reads like a sourced claim over a passage that reads like an unsupported opinion.
Authoritative quotations raise visibility. Quoting a credible third-party source with attribution lifts the passage in the same way cited statistics do. The engine’s synthesis step prefers passages that attribute, because the engine itself is trying to produce an answer that attributes.
The practical implication: every page targeting AI citation should contain at least two or three passages with cited data points or attributed quotations. These are the phrases the engine is most likely to pull into its answer.
This is not keyword stuffing for AI. It is sourcing discipline. The same editorial standards that make content credible to a human reader make it extractable to an AI engine.
Step 5: Make Sure Engines Can Fetch Your Pages
Content restructuring is wasted if the engine cannot reach the page. The technical prerequisites are specific and mostly small.
Server-side render critical content. Googlebot executes JavaScript reasonably well. Most AI-specific crawlers do not. Content that only appears after a React or Vue app boots is content the AI engine cannot see. Server-side rendering, static site generation, or a markdown-first publishing workflow ensures the content is in the HTML the crawler receives on the first request.
Allow AI crawlers in robots.txt. Five user agents need an explicit decision: GPTBot (OpenAI’s training crawler), OAI-SearchBot (OpenAI’s search crawler, a separate agent), ClaudeBot (Anthropic), PerplexityBot (Perplexity), and the Google-Extended control token. Blocking these by default means the engines they feed cannot see the content. The trade-off between training access and search access is real, but it should be a deliberate decision, not an accident of a copy-pasted template.
Publish an llm.txt file. A machine-readable content index at the site root, written for AI assistants. The standard is young and engine-specific, but the cost is low and the signal is clear. The llm.txt standard explained covers what the file is. The step-by-step guide to publishing an llm.txt file walks through implementation.
Add structured data. Article, FAQPage, and HowTo schema give engines a clean reading of what the page contains. Classical SEO hygiene that carries directly into GEO without modification. The companion post on Yoast SEO fields that carry into the AI era covers the metadata layer in detail.
None of these are large projects. A site that already runs a clean technical SEO program likely needs only the AI bot handling and llm.txt additions.
Step 6: Verify Across Engines
After restructuring and publishing, verify the work by repeating the audit from Step 1.
Prompt the same target queries into ChatGPT Search, Perplexity, and Google AI Overviews. Log whether the site now appears as a cited source. Run the same queries multiple times across a week, because citation is probabilistic. A single prompt returning a citation is encouraging. The same citation appearing across three or four runs of the same query is a confirmed win.
Three things to watch for.
Engine-specific results. A page can be cited by Perplexity but not by ChatGPT Search, or the reverse. This is normal. Each engine retrieves from a different index, ranks with a different model, and has different citation behavior. Perplexity tends to cite more aggressively and picks up content changes faster. ChatGPT Search leans on Bing’s index and favors established authority. Google AI Overviews uses Google’s own index and appears mainly on informational queries.
Citation position. Being cited as the first source in an answer is different from being cited as the fourth. Watch whether the site moves up in citation order over time. This is the GEO equivalent of moving from position five to position two in classical search.
Which passage gets extracted. The engine will sometimes cite a page but extract a different passage than expected. This is diagnostic. If the engine consistently pulls the third paragraph instead of the first, the passage structure of the opening needs work.
What Works Differently Across Engines
Each engine has quirks worth knowing before interpreting results.
ChatGPT Search retrieves through its Bing partnership. It favors pages with established authority in Bing’s index and tends to cite pages that have been indexed for longer. Inline citations appear as numbered footnotes. Response quality varies between the free and paid tiers. For content teams, the implication is that a page needs to be indexed in Bing (not just Google) to have a chance at ChatGPT Search citation.
Perplexity combines its own index with Google results and targeted real-time browsing. It cites more aggressively than any other engine, often listing five or six sources per answer. It picks up content changes faster than ChatGPT Search, which makes it the best engine for testing whether a restructured page is working. Perplexity is also the engine where passage structure matters most visibly, because its answers are heavily extractive.
Google AI Overviews uses Google’s own index and appears at the top of search results on a growing share of informational queries. Citation behavior is less consistent than Perplexity or ChatGPT Search. Some AI Overviews show clear source links. Some show none. The visibility gain from appearing in an AI Overview is significant because the overview sits above the organic results the user was going to scan anyway.
Gemini, Claude, and Copilot each answer questions from the web in certain contexts. Citation behavior varies. For most content teams, the three engines above are the ones worth tracking regularly. The others are worth a periodic check.
How Long This Takes
Honest timeline expectations for a team starting the process today.
Perplexity picks up content changes fastest. A restructured page can appear in Perplexity citations within one to two weeks of the content being re-indexed.
ChatGPT Search relies on Bing’s index, which recrawls less frequently than Google for most sites. Expect two to four weeks before restructured content shows up in ChatGPT Search citations.
Google AI Overviews uses Google’s own index, which is generally fast for pages Google already knows. Two to four weeks is a reasonable expectation for pages that were already indexed before the restructuring.
The feedback loop is slower than watching a rank tracker update daily. It is faster than classical link-building campaigns that take months to move a position. Weekly five-query audits are enough to see whether the work is producing results.
The compounding effect matters. A site that restructures five pages and sees citation-share gains on three of them has validated the approach for the entire site. The next twenty pages follow the same playbook, and the topical authority built by the first five makes the next twenty more likely to be cited.
Start This Week
Three actions, each targeting a different layer of the stack.
Monday: Run the five-query audit. Pick five questions a target customer would ask a chatbot. Prompt each into ChatGPT Search, Perplexity, and Google AI Overviews. Log which sources get cited. This is the baseline everything else is measured against.
Tuesday: Restructure one page. Pick the highest-value page on the site. Move the direct answer to the first paragraph. Add two cited statistics or authoritative quotations. Break long paragraphs into single-idea units. Use lists where enumerations exist in prose.
Wednesday: Check the technical floor. Verify that GPTBot, OAI-SearchBot, ClaudeBot, and PerplexityBot are allowed in robots.txt. If no llm.txt file exists, publish one. If critical content is JavaScript-rendered, flag it for server-side rendering.
Then wait two weeks and repeat the five-query audit. The difference between the first audit and the second is the first signal of whether the work is moving.
Clean markdown is the format AI answer engines prefer to read, which is why content teams building a GEO workflow often start by converting their research and source material into markdown before drafting. CopyMarkdown is the tool that makes that conversion a single click, from any webpage to clean markdown ready for an AI-first publishing pipeline.