How to Get Cited by AI Search Engines: A GEO Playbook for ChatGPT, Perplexity, AI Overviews and Copilot

Generative Engine Optimization (GEO) is the work of getting your pages quoted, and your brand named, inside AI-generated answers. It sits on top of SEO rather than replacing it: Google states that there are no additional requirements or special optimizations to appear in AI Overviews or AI Mode, and that a page must be indexed and eligible to show with a snippet. What changes is the unit that wins. An answer engine quotes passages, assembles shortlists, carries context across follow-up questions and leans on third-party opinion when it recommends something. This playbook covers each of those, then gives two procedures for finding where competitors are cited and you are not.

How AI answers choose what to cite

Most AI search products work as retrieval-augmented generation: they run one or more searches, read the retrieved pages, pick the passages that answer the question and write an answer with links to the sources they used. Google documents that AI Overviews and AI Mode may use query fan-out, issuing several related searches across subtopics to build one response. The retrieval, reranking and passage-selection mechanics, and how to structure pages so individual chunks survive them, are covered in the companion pillar on passage-level optimization. For this playbook, three consequences matter:

• Retrieval comes first. A page that is not indexed, or that blocks the engine's search crawler, cannot be cited. Which user agents to allow is its own decision, covered in the AI crawler access guide.
• The passage is the unit. The engine links a page, but it chose the page because one or two passages answered the question cleanly. Diffuse coverage of ten subtopics loses to one tight section on the exact question.
• Your own pages are one voice among several. For anything that sounds like a buying decision, the engine has good reason to discount what a brand says about itself and look for independent agreement.

There is one controlled study worth knowing. In GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024), researchers rewrote source pages and measured how much of each source appeared in generated answers on a 10,000-query benchmark. Adding citations to sources, quotations and statistics performed best, with a reported 30–40% relative improvement on their visibility metric; keyword stuffing did not help and on Perplexity performed worse than the unedited baseline. It is a lab setup with its own metric, not a promise for your site, but the direction matches everything below: verifiable specifics beat repetition.

Write answer-first pages

An answer-first section opens with the answer to the question its heading asks, then gives the reasoning, the exceptions and the detail. It is the same structure a good reference page has always had; AI retrieval just punishes its absence harder, because a passage that only makes sense after three paragraphs of setup is rarely the one lifted.

• Question-shaped headings. Write the heading as the question someone types: “How much does X cost?” rather than “Pricing overview.”
• A self-contained first sentence. It should answer the heading without the surrounding page: name the product or concept (not “it”), state the fact, give the condition. “X costs $29 a month on the Team plan, which includes…” works on its own; “As mentioned above, it's affordable” does not.
• Specifics you can stand behind. Real prices, named products, named standards, dates, and numbers with a linked source. Do not invent statistics to look citable; a wrong number that gets quoted is a liability with your name on it.
• Tables and numbered steps where the content is a comparison or a process. They map cleanly onto the lists and tables AI answers render.
• Honest dates. Show a real “updated” date and update the content when it changes. Bumping a timestamp without changing anything helps no one; a refresh routine is in the content refresh calendar guide.
• Links to definitions and deeper pages. The first mention of a term should link to the page that defines it. It helps readers and gives an engine that follows links more of your context.

FAQ blocks that get cited

A good FAQ block is a stack of short, self-contained question-and-answer passages, which is exactly the shape answer engines lift. A bad one is an accordion of support boilerplate nobody searches for. The difference is where the questions come from and how tightly the answers are written.

Mine real questions, then cluster them

1. Support tickets and chat logs. The questions your team answers every week are the ones buyers ask AI tools.
2. Sales-call notes and transcripts. Richer than tickets because they contain comparison and objection questions (“What about…”, “Can you…”).
3. Search Console queries. Filter for what/how/why/can/does/is queries your pages already get impressions for.
4. People Also Ask and your AI-visibility prompt set. The questions search already associates with your topic.
5. Reddit, Quora and niche forums. Buyer questions in the buyer's own words.

Group the questions by intent: definition, comparison, how-to, troubleshooting, eligibility, pricing, integration. Each cluster belongs on the page that matches it: definitions on a glossary entry, comparisons on the comparison page, how-to questions on the guide. One mega-FAQ for the whole site helps nobody.

Question shape

• Use the literal phrasing people type, starting with what, how, why, can, does or is.
• Be specific: “How much does X cost per month?” beats “Pricing.”
• Use the buyer's vocabulary in the question; you can introduce your preferred term in the answer.
• No wordplay. “Will AI UGC replace human creators?” is findable; “Is AI UGC the death of creators?” is not.
• One question per entry. Split “What is X and how do I set it up?” into two.

Answer shape

• One short paragraph. The first sentence fully answers the question; the rest adds the condition, the exception or the next step.
• No openers (“Great question”), no “it depends” without saying on what. Answer the common case and name the exception in one sentence.
• Named entities and sourced numbers instead of “most platforms” and “many brands.”
• One to three links to the glossary entry or guide that goes deeper, with descriptive anchor text.
• Read it aloud as if a voice assistant were saying it. If it rambles, cut it.

FAQPage markup: what it does and what it doesn't

Structured data is a clean, machine-readable copy of the questions and answers, and it is cheap to add. Two facts keep expectations honest. Google says there is no special schema needed to appear in its AI features. And since 2023 Google shows FAQ rich results only for well-known, authoritative government and health websites. So add FAQPage markup because it describes the page accurately, not because it unlocks anything on its own. The rules that matter:

• Visible text must match the markup. Render the questions and answers as real HTML (headings and paragraphs, or an accordion whose content is in the server-rendered HTML) and emit the same text in the JSON-LD. Google lists matching structured data to visible text as a basic practice.
• Plain text in the markup. Links and emphasis can live in the visible version.
• One FAQPage block per page, combining every question into a single mainEntity array.
• Validate with the Schema Markup Validator or Google's Rich Results Test, then run the questions through the engines in a fresh session.

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "How much does X cost per month?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "X costs $29 a month on the Team plan, which includes..."
      }
    }
  ]
}

Where FAQ blocks belong, and the mistakes to avoid

• Long-form guides and glossary entries: at the end, for the long-tail questions the body did not answer.
• Pricing, comparison and product pages: near the call to action, for objections and eligibility questions.
• Use-case and location pages: a topical FAQ that is genuinely different from sibling pages, not the same five questions everywhere.
• Skip it on checkout pages, thin collection pages and anything where the questions would be invented.
• Common failures: return-policy questions on a topic page (move them to the help center), compound questions, hedged answers, markup without visible text or the reverse, FAQ content loaded only by client-side JavaScript, and answers with years-old facts that nobody revisits.

Comparison queries: getting into “best X” and “X vs Y” answers

“Best project-management tool for small teams”, “Notion vs Obsidian”, “alternatives to X”: these are where purchase decisions happen, and the engine is not quoting one source. It is building a shortlist. A useful working model (ours, not a documented engine spec) has four stages:

1. Candidates. Products that show up alongside the category and its qualifier across listicles, reviews, forums, comparison pages and the model's training data.
2. Filter. Candidates with thin, stale or one-source evidence drop out.
3. Rank and describe. Each survivor gets a position and a one-line “best for…” description drawn from what sources say about it.
4. Render. A list or table, with citations on the claims about each entry.

The consideration set is qualifier-specific: “best CRM for small teams” and “best CRM for enterprise” produce different shortlists. A brand that never enters the candidate pool cannot be ranked however good its pages are. Entering it is a co-occurrence problem: be named with the category and the qualifier in places you do not control, and cover each qualifier you want to be eligible for.

Build comparison pages an engine can quote

• A real table with named attributes (price, best for, main limitation, key integrations), not marketing adjectives.
• An honest “best for” line for every option, competitors included. A page that trashes every alternative reads as an ad.
• Explicit qualifier coverage. If “for small teams” and “for enterprise” are separate searches, address both.
• Current, dated prices and limits instead of “flexible plans.”
• Product imagery that matches the claim. If you are “best for outdoor use”, show the product outdoors; answers that render thumbnails pull them from pages associated with you.

Our own comparison pages follow this pattern. When an answer describes you wrongly (“the budget option” when you are premium), the fix is at the source: state the positioning plainly on your pages and get independent sources to echo it.

Follow-up questions: staying cited across a conversation

AI search is conversational. After the first answer, users refine (“what about under $100?”), compare (“how does that compare to X?”), clarify (“does it work with Shopify?”) or pivot (“what about for B2B?”). The engine answers those with the earlier conversation as context, often running new searches. A brand cited in the first answer can disappear by the third. The engines do not publish how they weight earlier turns, so treat the following as working heuristics and test them:

• Name the entity in each section's first sentence. A passage that says “it integrates with Shopify” depends on context the follow-up search may not retrieve. A passage that says “X integrates with Shopify through…” stands alone.
• Anchor claims on the use case, not one query. “X wins the Y comparison” dies when the user pivots. “X is built for small agencies that need Z” still applies to the follow-ups.
• Cover the predictable next question. For each priority query, list the four follow-up types above and make sure each has an answer somewhere on your site: a section, an FAQ entry or a sibling page for the adjacent buyer or use case.
• Keep the brand visually consistent. If answers show images, the same product shots and the same on-camera persona across your pages make it obvious they belong to one brand.

To test it, write three-to-five-turn scripts for your priority queries (one head question followed by the follow-ups you see in sales calls and support), run them in each engine, and log for every turn whether you are cited, paraphrased, merely listed or gone. Where you drop, the pattern usually points to the fix: lost on a clarification means the answer is missing from your site; lost on a pivot means the adjacent page does not exist; replaced by a competitor saying the same thing means your passage did not name you.

Citation is not recommendation: consensus and sentiment

Read real answers to commercial questions and three outcomes appear. Cited, not recommended: the engine quotes your page for a spec, then recommends a competitor. Recommended, not cited: the engine names you as the pick based on a forum thread or a review roundup and never links your site, which is why share of voice has to count mentions, not only links. Neither: the evidence is thin or mixed and the answer hedges or routes to the brands it has more agreement on.

The off-site evidence an engine can lean on, which this site's glossary calls the consensus signal, comes from three places:

• Community discussion: Reddit, niche forums, Q&A sites. Reads as unmotivated peer experience, and one well-ranked “what's the best X” thread can shape how you are described.
• Reviews: volume, recency and sentiment on marketplaces, app stores and review sites in your category.
• Independent editorial: “best X” roundups, comparison articles and expert reviews.

Volume amplifies whatever sentiment you already have. A small, specific, positive body of reviews (“I use this for X and it works because Y”) maps directly onto “best X for Y” questions. A large body with a loud recurring complaint gets that complaint repeated in the answer. A body that went quiet two years ago describes an old product. The target is recent, specific and net-positive, with enough volume not to look anecdotal.

The consensus audit

1. Run your commercial prompts (“best X”, “X vs Y”, “is X worth it”) in each engine and record whether you are named, cited or recommended, and which off-site sources the answer references.
2. Find the top-ranking Reddit and forum threads for your category and note how and when your brand appears, or that it does not.
3. Score volume, recency and sentiment on the two or three review sites the answers actually reference, against your top competitors.
4. List the “best X” roundups that rank, whether you are in them and in what position.
5. For each surface, rate yourself against the strongest competitor. The biggest gaps on the surfaces the answers cite most are the backlog.

Legitimate influence only

You cannot fake this, and trying is illegal in the US as well as self-defeating. The FTC's Consumer Reviews and Testimonials Rule, in effect since October 21, 2024, bans writing or buying fake reviews, requires insider reviews to disclose the relationship, and bars incentives that require a review to be positive. Community platforms also remove astroturfing. What works: fix the complaints that keep showing up, ask real customers for specific reviews after purchase, participate openly as the brand where a community allows it, give creators and reviewers product access with disclosure, and earn roundup inclusion on merit. You optimize the product and the evidence trail, and the consensus follows.

Procedure: map competitors' citation footprints

A citation footprint is the inventory of a competitor's URLs that AI answers cite, with the prompt that triggered each citation and what the answer said. It shows which pages you should have and do not.

1. Fix one prompt set. Use the same commercial and comparison prompts you track for your own brand, so the two views are comparable. Change it quarterly, not weekly.
2. Pick two or three competitors. Run the prompt set and tally which brands appear most. Prefer brands that show up consistently in at least two engines over one that dominates a single engine. Adjacent-category leaders often teach more than direct clones. Mapping ten competitors makes the data thin everywhere.
3. Capture in a flat sheet with the columns below. A spreadsheet the writer can annotate beats a dashboard nobody reads.
4. Start with two engines, then add others one at a time, in order of where your buyers actually search (Amazon's Rufus if Amazon is a channel, for example).
5. Ignore one-off citations. Answers vary between runs. Only treat a competitor page as part of the footprint once it has been cited in two consecutive capture rounds.

Procedure: turn the map into a content gap audit

1. Pivot by page type. Count distinct cited URLs per page type per competitor. The type where they have many cited pages and you have none is usually the biggest single investment: a missing set of comparison pages, use-case pages or a glossary cluster.
2. Pivot by prompt. Score each competitor and yourself per prompt on a simple scale (0 not cited, 1 cited low in the answer, 2 near the top, 3 first). Cells where they score 2–3 and you score 0–1 are the gaps. Order them by how important the prompt is to the category; five head prompts outweigh twenty long-tail ones.
3. Cluster what the answers say. Group the captured sentences by claim type: use case (“best for X”), comparison (“cheaper than Y”), social proof (“reviewers say”), specification, visual. The mix tells you what kind of statement the engines treat as an answer in your category. If they quote competitors' use-case lines and your pages are all specs, you have been writing the wrong kind of sentence.
4. Write one brief per page type and prompt cluster, not one per long-tail prompt, which produces thin pages. Quote the competitor's cited sentence verbatim in the brief so the writer knows what claim the new page must make better, with your own evidence.
5. Attach a visual brief when the answers for those prompts show images: what the product should be shown doing, and in what setting.
6. Cap the backlog at what the team can publish well each week, publish, and recapture. The next capture round tells you whether the new page entered the answers. Give changes several weeks before judging them; engines recrawl on their own schedules.

One cheap extra read: if a competitor publishes an llms.txt file, it lists the pages they consider most important. Diffing it occasionally shows new pages as they launch. Whether any engine uses the file is unsettled; the llms.txt implementation guide covers what it is and how to publish your own.

Measure citations and recommendations separately

AI answers vary between runs, accounts and locations, so treat any single screenshot as an anecdote. Keep a fixed prompt set, run each prompt several times per engine on a schedule, and log:

• Named, cited and recommended, as three separate fields.
• Position in the shortlist and the “best for” description the answer gave you, and whether it matches your intended positioning.
• For scripted conversations, the same fields per turn.
• Which of your URLs, and which third-party URLs, were cited.
• Factual errors about your brand. Wrong facts usually trace back to an outdated or ambiguous source; fixing your entity data is covered in brand entity disambiguation.

Tools such as Profound, Otterly.AI, Peec AI and AthenaHQ automate the prompt runs and competitor breakdowns; a spreadsheet and a weekly manual panel work fine at small scale. To connect citations to traffic and revenue, see AI search attribution tracking.

A sensible order of work

1. Confirm the priority pages are indexed and crawlable by the engines you care about.
2. Fix the prompt set and take a baseline: your brand and two or three competitors.
3. Rewrite the top pages answer-first and add FAQs built from real questions.
4. Publish or fix the comparison and use-case pages the gap audit says are missing.
5. Run the consensus audit and start on reviews and the recurring complaints.
6. Script follow-up conversations for the head queries and fill the missing answers.
7. Recapture on a schedule and repeat.

Where product photos and video fit

Engines cite text. Images matter in two narrower ways: some answers render thumbnails or image results next to shortlisted products, and a guide or comparison page with real product shots is more useful than one without. The rule from the comparison section applies: show the product in the setting your positioning claims, and keep the look and on-camera persona consistent across pages. Video answers and image carousels are covered in video answer optimization.

If you need those visuals at volume, ppl.studio generates UGC-style product photos and short videos with your product as the prop and a reusable AI Expert as the consistent face. Do not use generated people as fake customers in reviews or testimonials; that is the kind of fabricated endorsement the FTC rule above prohibits. For physical locations, the local version of this playbook is local AI search optimization; for shopping assistants such as Rufus, see the AI shopping assistants playbook.

Frequently Asked Questions