Answer Engine Optimisation: the 2026 playbook for ecommerce

That second figure is the case for the conversational-PDP lever above in one number: an agent has already filtered for the shopper before the click, so it sends them straight to the product, not the homepage. A PDP that cannot stand alone, no category context, no "who it is for" line, loses a shopper who was never going to see anything else first.

Is AEO just SEO with new vocabulary?

The temptation, if you have run an SEO programme for ten years, is to map AEO onto your existing mental model: keywords become "topics agents care about", backlinks become "brand mentions on credible sites", rankings become "citation positions inside engines". Some of that is right. Most of it is wrong, because the underlying mechanism is different.

In classical SEO, you optimise to rank a page on a SERP, and the user clicks (or does not). In AEO, you optimise to be quoted inside the agent’s answer, sometimes with a link, often without. The agent has already read your page. The question is whether what it read was good enough to surface. The teardowns in how AI agents actually choose products show exactly which sentences make the cut.

That changes everything downstream. Page authority matters less; page legibility to a language model matters more. Anchor text matters less; structured evidence on the page matters more. Keyword density matters not at all; JSON-LD coverage matters enormously. AEO is closer to writing a press release a journalist will read than to ranking a page a crawler will index.

Which four levers actually move citation rate?

Across the brands we have audited, the same four levers explain the bulk of the variance in citation rate. They are not the only things that matter, but they are where 80% of the lift comes from, and the order below is the order of ROI.

1. JSON-LD coverage, and the shapes that matter

AI engines lean on schema.org JSON-LD because it is the only structured contract between a page and a model that has had a decade to mature. The shapes that have measurably moved citation rate in our testing: Product (price, availability, GTIN), Review + AggregateRating, FAQPage, HowTo, BreadcrumbList, Organization, WebSite with SearchAction. Article-level: Article + Person author with a real url resolving to that person.

The mistake to avoid: rich-results-test-pass-rate is not the same as citation rate. Some shapes (Speakable, ImageObject) pass validation but agents do not seem to use them as evidence. Some shapes (Review) are heavily weighted by agents but easily faked, so spam patterns get demoted aggressively. Focus on the shapes the engines have explicitly named in their published agent docs (OpenAI, Anthropic, Perplexity, Google).

2. llms.txt, the file you can ship this afternoon

llms.txt is robots.txt for language models. A single markdown file at /llms.txt that tells an agent which parts of your site it can read, where the catalogue lives, where the authoritative product pages are, and what canonical URLs to cite. As of mid-2026, only 21% of brands ship one. The ones that do see a measurable lift in citation accuracy (fewer hallucinated facts about their products) within weeks.

The minimum viable llms.txt for an ecommerce brand is six sections: About, Products (link to the product feed), Reviews (link to the structured review index), Policies (returns, shipping, warranty), Citation guidance (which URLs you want quoted), and a Contact path for agent owners who hit errors. Ten minutes of writing, days of value.

3. Conversational PDP structure, the seven-second window

Agents read your PDP the way a shopper does, top to bottom, with a hard attention cliff after the first two paragraphs. If your PDP buries the answer to "what is this and who is it for" past a hero image and a 12-line spec block, you have already lost the agent. Restructure: first paragraph is the answer. Second paragraph is the differentiator. Then specs, then reviews, then social proof. The full mechanic is in the 7-second window.

PDPs that are easy for an engine to quote tend to share three structural traits: a one-sentence "what it is" lead, an explicit "who it is for" line, and a "what makes it different" tile. None of these things require redesigning the page, they are usually a CMS field away. Most teams ship them in a sprint.

4. Verified-review evidence layer

Agents weight evidence by its falsifiability. A press quote on the brand’s own page is weak, anyone can write one. A verified-purchase review with a timestamp and a rating distribution and a photo is strong, the agent can reason about it. Surfacing verified review media inline, with Review markup an engine can parse, gives it evidence it can reason about; a text-only review count does not.

The cheap way to do this if you already run a review platform: pull the verified-review JSON-LD off the platform’s API and surface it inline on the PDP, not behind a reviews tab. The expensive way: rebuild your reviews band from scratch. Most teams should do the cheap version this quarter and the expensive version next year.

Citation rate is the new ranking. Track it that way.

You did keyword rank tracking in 2014; now you do citation rate tracking. The mechanic: a panel of 100–300 prompts that represent how shoppers in your category actually ask AI engines for product help, run weekly across ChatGPT, Claude, Perplexity, Gemini, and recorded into a citation log.

What you measure: citation rate (prompts where you appear), citation position (first cited / fifth cited / last cited), citation accuracy (do the facts the agent quoted match the source), and competitive citation share (out of all brands cited on a prompt, what % are you).

Track this in a flat spreadsheet for the first quarter. Move to proper tooling (Profound, AthenaHQ, Otterly, or build something internal) once the practice is real.

What to actively deprioritise

The hardest part of an AEO programme is letting go of work that used to matter. Three things have measurably stopped paying back in our testing, keep doing them at maintenance level, but stop budgeting fresh hours against them.

1. 1Pure backlink building, agents cite credible sources, but their definition of "credible" is closer to "well-structured + factually consistent" than to "DR 60+". Stop buying links. Save the budget for evidence.
2. 2Keyword density grinding, agents read pages as semantic units, not as keyword bags. Density above natural is now a demotion signal, not a boost.
3. 3Content velocity for its own sake, shipping ten mediocre blog posts a week to feed the keyword spreadsheet is now actively harmful. Ship one good piece a fortnight. Make it citable.

A 90-day AEO programme

1. 1Week 1 · Citation baseline. Run your top 100 category prompts across the four major engines. Record citation rate. This is your starting line.
2. 2Weeks 2–3 · JSON-LD audit + remediation. Run every top-traffic PDP through the Rich Results Test. Add the six required Product fields where missing. Add Review + AggregateRating if you have a review platform. Add FAQPage on pages with Q&A.
3. 3Week 4 · llms.txt. Ship the six-section file. Test by asking ChatGPT and Perplexity to look up your brand "as a shopper would". Iterate the file based on what the agents return.
4. 4Weeks 5–8 · Conversational PDP restructure. First paragraph = answer. Second = differentiator. Third = who it is for. Reviews-band inline (not tabbed). Verified-purchase badges prominent.
5. 5Weeks 9–10 · Evidence layer. Pull verified-review JSON-LD inline. Map review photos / videos onto the PDP gallery alongside curated UGC. Surface helpful-vote chips visibly.
6. 6Weeks 11–12 · Citation re-baseline. Re-run the 100 prompts. Compare. Iterate the lowest-citation prompts. Write up the playbook for the next 100.

The biggest mistakes brands make in their first AEO programme

• Treating AEO as a new content stream. It is not, it is structural work on the pages you already have. Most of the wins are within your existing PDP / category page / FAQ surfaces.
• Optimising for one engine. ChatGPT, Claude, Perplexity, Gemini all read pages slightly differently. Optimising for one without monitoring the others will silently rank-tank you on the others.
• Faking the reviews JSON-LD. Engines have gotten very good at detecting unverified-purchase review claims. The penalty is permanent demotion, not a tap on the wrist. Only emit Review + AggregateRating JSON-LD for reviews you can prove are real.
• Skipping the citation log. Without weekly tracking, the team will pivot on vibes. With it, the programme stays disciplined.
• Hiring a separate AEO team. The team that runs your SEO + content + structured data should run AEO. New skills, same team.

Your action list, this week

1. 1Run 25 representative prompts across ChatGPT + Perplexity + Gemini. Record whether each cites you, and where you sit in the cited list. That is your baseline.
2. 2Ship llms.txt. Six sections, 200 lines, ten minutes.
3. 3Pick your top-traffic PDP. Audit it for JSON-LD coverage against the six required Product fields + Review + AggregateRating + FAQPage. Fix everything red.
4. 4Restructure the first 100 words of that same PDP to lead with the answer, the differentiator and the buyer profile. Ship it as a single PR.
5. 5Schedule a weekly half-hour to re-run the 25 prompts and log citation movement. This is the new keyword review.