The Answer Engine Optimisation (AEO) Playbook for 2026: What Actually Moves Citations

Answer engine optimisation means restructuring your site and content so ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews can retrieve, extract and cite you in the answers they generate. In practice that's six things done in sequence over roughly 90 days: map what buyers actually ask, measure your starting point, fix the technical plumbing, ship pages engines retrieve, earn off-site citations, then refresh and re-measure. Each step below covers what to do, how to check it worked, and where teams usually go wrong.

We won't re-run the AEO vs GEO vs SEO definitions debate here; that's covered properly in our explainer on the differences. This is the execution sequence.

Why sequence the work over 90 days instead of doing it all at once?

Because each step depends on the one before it. You can't measure a baseline before you know which prompts matter, you can't judge whether a new page worked before engines have re-crawled it, and you can't publish confidently until the site's plumbing stops silently blocking the bots that would retrieve it.

Here's the full sequence at a glance.

Phases overlap deliberately. Off-site citation work should start as soon as the first owned pages are live, not after the whole build phase finishes.

Step 1: Map buyer prompts

What to do. List the actual questions your buyers ask AI engines before they ever reach your site: comparison questions ("X vs Y"), evaluative questions ("is X worth it"), and pricing or fit questions ("how much does X cost", "is X good for [use case]"). Pull these from sales call transcripts, support tickets, your existing keyword data and a manual pass typing real buyer language into ChatGPT and Perplexity yourself. Aim for 40 to 100 prompts, tagged by funnel stage (awareness, comparison, decision) and by whether they're branded or unbranded.

How to verify it worked. Run five of your mapped prompts through ChatGPT, Perplexity and Google AI Overviews by hand. If the answers already reference companies, products or content types that look like your market, the prompt list is realistic. If the answers feel generic or off-target, the prompts are too broad or copied from generic keyword tools rather than real buyer language.

Common failure mode. Reusing an old SEO keyword list unchanged. AI prompts run roughly five times longer than search keywords and carry more clauses and context, so a keyword like "CRM software" needs to become something closer to "which CRM works best for a 20-person sales team that's outgrown spreadsheets." Short keywords under-represent what buyers actually type into a chat box.

Step 2: Baseline measurement

What to do. Before changing anything, record where you stand. Track two separate numbers per engine: visibility (the share of tracked answers that mention your brand at all) and share of voice (your share of brand mentions among answers that mention any brand in your category). These measure different things. Also record which third-party domains get cited most often for your prompt set; that tells you where off-site work should focus later.

How to verify it worked. You should end this step with a dated snapshot: visibility and share of voice per engine, plus a short list of the domains currently winning citations in your category. Our own 30-day tracking of AI engine citations across five engines (June 5 to July 5, 2026) shows exactly why per-engine detail matters: across 603 answers, Schmitdy's visibility ran 6.15% on Claude, 6.00% on ChatGPT, 5.54% on Gemini and 4.50% on Perplexity, and 0.00% on Google AI Overviews in that window. A brand can be genuinely present on four engines and structurally invisible on the fifth; an average across engines would have hidden that gap completely.

Common failure mode. Averaging across engines, or measuring only one. The same window also showed Schmitdy ranking 5th of 13 tracked brands on visibility (breadth: how often it's mentioned at all) while ranking 2nd on share of voice at 15.2% (depth: how dominant the mention is when it happens). Breadth and depth are different problems needing different fixes, and a single blended score erases the distinction that tells you which one to work on.

Step 3: Entity and site plumbing

What to do. Three things, roughly in this order:

First, confirm AI crawlers can actually reach your pages. AI bots now split into two categories with different jobs: training crawlers like GPTBot and ClaudeBot, and retrieval or search crawlers like OAI-SearchBot, Claude-SearchBot and PerplexityBot. You can block the training crawlers on copyright or competitive grounds and still stay fully eligible for AI-search citations by explicitly allowing the search-oriented ones in robots.txt.

Second, fix structured data. Mark up FAQ content, product data and organisation details with schema.org types, using Google's structured data guidelines as the reference for correct implementation. Be clear-eyed about what this buys you: Google stopped showing FAQ rich results in Search from 7 May 2026 and dropped HowTo rich results back in 2023, so neither produces a visible Google snippet any more. The markup still matters because it helps AI engines parse entities and structure on the page; treat it as a machine-readability aid, not a SERP feature.

Third, restructure pages to answer-first: a direct 30 to 60 word answer at the top of the page, before any brand story or navigation, followed by short atomic paragraphs engines can lift as standalone chunks.

An llms.txt file (a plain Markdown file at yourdomain.com/llms.txt listing key pages) is a genuinely low-cost addition worth doing, but keep expectations honest: it remains a proposed convention, and no major AI lab has confirmed its crawlers act on it, so it's a nice-to-have layered on top of real crawler access, not a substitute for it.

How to verify it worked. Check your server logs or a bot-analytics tool for confirmed visits from OAI-SearchBot, Claude-SearchBot and PerplexityBot within two weeks of opening access. Run your key pages through Google's Rich Results Test to confirm schema validates without errors (even though FAQ and HowTo won't render as a snippet, valid markup confirms the structure parses cleanly). Read your top three pages cold and check whether the first 60 words answer the implied question without needing the rest of the page for context.

Common failure mode. Blocking all AI bots by default out of a training/copyright concern, without carving out the search-retrieval crawlers separately. That single misconfiguration removes you from AI-search citation eligibility entirely, regardless of how good the content behind it is.

Step 4: Owned content that earns retrieval

What to do. Publish or rewrite pages in the shapes engines actually retrieve for your category, not the shapes that feel safest to write. Alternatives and comparison pages, ranked listicles with named criteria, and direct question-and-answer content outperform generic explainer prose by a wide margin. Every page needs a genuine reason to be cited: a stat you measured yourself, a named framework, or a comparison table nobody else has built honestly.

The page-shape gap shows up clearly in our own tracking of which pages engines retrieve and cite in our category:

How to verify it worked. Two to four weeks after publishing, run your target prompts again and check whether the new page (or its specific sections) shows up as a cited source, not just whether your brand gets mentioned in passing. Watch server logs for repeat visits from search-retrieval bots to the new URL; a page that's actually being used for answers gets re-crawled, not visited once.

Common failure mode. Publishing a self-ranking listicle that puts your own product first with no real comparison. Engines and readers both discount transparently self-serving rankings, and it's reputationally weak even when it works short-term. If you cover your own product in a comparison, disclose it plainly (as we do throughout this site) rather than pretending the ranking is neutral.

Step 5: Off-site citations

What to do. Engines don't just cite your own site; they cite wherever your category gets discussed, and that's disproportionately a small set of third-party domains. Our own 30-day tracking of AI engine citations across five engines (June 5 to July 5, 2026) found Reddit was by far the most-retrieved third-party domain for AI-search-category prompts, at 276 retrievals, ahead of LinkedIn (140), Semrush (93), Search Engine Land (77), Medium (70) and G2 (51). This isn't unique to our category: an independent Search Engine Land study found Reddit, YouTube and LinkedIn are the most-cited sources across ChatGPT, Perplexity and Google AI Overviews generally.

Prioritise, in order: genuine participation in Reddit threads where your category gets discussed (answer the actual question first, mention your product only if it fits naturally), placement in third-party "best X" listicles and comparison roundups, earned editorial coverage from publications AI engines already cite in your space, and complete, current profiles on review sites like G2 where your category gets reviewed.

How to verify it worked. Track citation-source data per engine monthly and watch whether previously absent domains (Reddit threads mentioning you, a listicle that added you, a new G2 review) start appearing as cited sources for your mapped prompts. A single new mention rarely moves the needle; look for the trend across a full month.

Common failure mode. Posting obviously promotional content to Reddit or a forum. Threads read as authentic get retrieved and trusted; threads that read as thinly-veiled marketing get downvoted, removed, or simply ignored by both readers and the retrieval systems that increasingly weight engagement signals.

Step 6: Refresh and re-measure

What to do. AI answers shift continuously as engines re-crawl and models update, so a one-off optimisation pass decays. Set a standing cadence: re-run your mapped prompts monthly, refresh your two highest-citation pages every quarter with updated numbers and examples, and do a full prompt-map refresh every six months as buyer language and the competitive set both shift.

How to verify it worked. Your visibility and share-of-voice numbers should be trending in a direction you can explain, not just moving randomly. If a page's citation count drops sharply, check first whether the underlying facts went stale (pricing, feature claims, statistics) before assuming the format stopped working.

Common failure mode. Treating the 90-day programme as a one-time project rather than a standing operating rhythm. The first 90 days establish the machinery; visibility earned in month three erodes by month six without a refresh cadence behind it.

What does this look like end to end?

Put together, the six steps form a loop rather than a straight line: map, measure, fix, build, earn, then measure again, feeding what you learn back into the next mapping pass. Teams that treat it as a checklist to finish once tend to see an initial lift that fades. Teams that treat step 6 as the start of a new loop tend to compound.

Running this entire sequence properly, especially steps 4 and 5, is a full-time job for someone. If you'd rather see where your brand currently stands across the five engines before committing a person or a quarter to it, our free AI search audit gives you the baseline numbers from step 2 without the setup work.