Getting your store cited by AI search — a field checklist

There is a lot of advice about optimising for AI search. Most of it is written by people who have not checked what these systems actually fetch.

We recently took a production storefront — a D2C brand on Shopify, headless (custom frontend tech stack), around 65 pages of products plus editorial — through the full answer-engine treatment. Before implementing anything we tested every recommended tactic against primary sources: crawler documentation, bot-log studies, other people's controlled experiments. Then we shipped what survived.

Verdicts first, receipts after.

The verdict table

The distinction everyone gets wrong

You will find one camp insisting schema is how you win AI search, and another citing hard evidence that it does nothing. Both are looking at real data. They are describing different pipelines.

Retrieval is an engine fetching your URL to answer a question. Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 controls: AI Mode moved +2.4% and ChatGPT +2.2%, neither statistically significant, and AI Overviews moved −4.6%, which was. A separate searchVIU experiment watched ChatGPT, Claude, Perplexity, Gemini and Google AI Mode fetch pages live to see what each one read. None used schema markup. Every system extracted visible HTML only.

Grounding is an engine answering from a structured index it already holds. Google's AI Mode grounds product answers in the Shopping Graph, and the Shopping Graph is fed by your Merchant Center feed and your Product markup — not by fetching your page mid-conversation.

So schema is close to worthless for getting an article cited, and load-bearing for getting a product surfaced. If you sell things, that is the whole argument, and it is why the advice sounds contradictory: people generalise from whichever pipeline they happened to measure.

The rest of this follows from that split.

Gate 1 — can a machine read you at all?

Binary checks. Failing any of these makes everything downstream irrelevant, which is why so much AI-search work produces nothing.

• Product pages render server-side. Fetch one as a crawler and count the words: curl -s -A "GPTBot/1.0" https://yourstore.com/products/x | wc -w. A near-empty shell means you are invisible to most of them. None of the third-party AI crawlers render JavaScript; Google's infrastructure does, which is how a JS-only store can appear in AI Overviews while ChatGPT and Perplexity see nothing at all. Either way the failure is silent — the page looks perfect in a browser.
• Test each crawler by name. GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, ClaudeBot, Google-Extended, Bingbot. Record the status code and word count for each.
• Write explicit allow-groups in robots.txt. AI crawlers come in three classes: training bots (GPTBot, ClaudeBot, Google-Extended, CCBot), search-index bots (OAI-SearchBot, PerplexityBot), and user-triggered fetchers (ChatGPT-User, Perplexity-User). Blocking the second removes you from AI answers outright — OpenAI's own documentation says blocking OAI-SearchBot removes you from ChatGPT search.
• Restate your disallows inside every named group. A named User-agent: group replaces the User-agent: * rules for that bot rather than adding to them. Miss it and the disallows you wrote once under User-agent: * simply do not exist for any bot you named — it will happily crawl /search, /cart and every faceted URL you meant to keep out of an index.
• Check your CDN is not challenging AI crawlers. A 403 to PerplexityBot is a decision someone made in a WAF rule, possibly by accident, and host-level default blocking is spreading.
• Canonicalisation is clean. http→https, www→non-www and trailing-slash variants all 301 to one form; missing pages return a real 404 rather than a soft 200.

Gate 2 — the feed is the biggest lever you have

For a store, your Merchant Center feed now matters more than your website for AI shopping visibility. This is the highest-return item on the list, and the one most stores have not touched.

In January 2026 Google added six conversational attributes, documented as helping "AI systems and conversational agents better understand your products' specific nuances," with AI Mode named as a target surface.

• question_and_answer — real customer questions and answers.
• document_link — manuals, spec sheets, care instructions.
• related_product — curated cross-sells, not algorithmic ones.
• item_group_title and variant_option — how variants relate.
• popularity_rank — relative demand within your catalogue.
• Derive every one from existing data. Nothing should be written twice. The FAQ set that renders on the page ships as the feed's Q&A pairs; drift between the two is worse than either alone.
• Establish the brand entity. Organization markup with sameAs pointing at your real profiles. Cheapest block on the page, and it is grounding work rather than citation work — it is what lets a system treat your brand as a disambiguated thing it can attribute claims to.
• Fill the ordinary attributes properly too — GTINs, short_title, product_highlight, product_detail, lifestyle_image_link. These predate the conversational set and still carry weight.

Gate 3 — is there anything worth citing?

Where stores actually lose. Gate 1 is plumbing with a known fix; this is a content problem, and it is the part worth paying for.

An assistant answering "waterproof walking boots under £200 that fit wide feet" needs facts it can verify. If your page says "engineered for the outdoors" and a competitor's says "Gore-Tex lined, available in EE and EEE widths, 480g per boot", the competitor gets cited — not because their SEO is better, but because yours contains nothing to cite.

Remember what retrieval reads: visible HTML. Everything here has to be on the page as text.

• Specifications as text, never as an image. A spec table exported as a PNG is unreadable to everything that matters.
• The unglamorous attributes. Dishwasher safe. Machine washable. Fits through a standard doorway. Exactly what buyers ask assistants, and exactly what marketing copy omits.
• Sizing in the buyer's words. "Runs half a size small" is what someone asks about; "true to form" is what a brand writes.
• Answer the question in the first two sentences, then elaborate. Extraction favours self-contained passages of roughly 40–80 words under a heading phrased as the question.
• Name the entity in prose rather than "we" and "our." A passage lifted out of context still has to say who it is about.
• Reviews and customer Q&A rendered server-side, not injected by a third-party widget after load.
• Facts consistent across every page that states them. Inconsistency reads as unreliability to a machine deciding whether to cite you.
• Stop treating rank as the route. AI Overview citations coming from top-10 organic results fell from 76% to 38% between July 2025 and March 2026. Ranking and being cited have decoupled.

Gate 4 — register where the retrieval happens

• Bing Webmaster Tools. ChatGPT Search retrieves from Bing's index, Copilot runs on it, DuckDuckGo feeds from it. Free, ten minutes, one-click import of your Search Console verification.
• IndexNow — a static key file plus a ping when URLs change. Updates reach that whole ecosystem in hours rather than crawl-schedule weeks. Google does not participate; that is fine, it is not for Google.
• Submit the child sitemaps, not just the index. A sitemap index discovers sitemaps, not pages, so submitting it alone leaves a dashboard reading "1 URL discovered" while everything looks correctly filed. On Shopify that means sitemap_products_1.xml and its siblings, not just /sitemap.xml.

What we deliberately skipped

• Schema as an AI-citation service. 1,885 pages, 4,000 controls, no causal uplift, −4.6% on AI Overviews, five engines confirmed to ignore JSON-LD on retrieval. Keep Product and Offer markup for the Shopping Graph and Google's rich results. Do not buy it as a citation play.
• Chasing FAQPage rich results. Google restricted these to government and health sites in August 2023 and removed them entirely in May 2026. The markup is still parsed, and still worth shipping if your shopping surfaces consume it — but it earns nothing in the SERP, and the visible Q&A on the page is what retrieval actually reads.
• llms.txt. Bot-log analysis across hundreds of millions of requests shows GPTBot, ClaudeBot and PerplexityBot essentially never fetch it, and Google compared it to the keywords meta tag. Its real users are developer-tool agents. Ship a minimal one if you like — an hour, and it helps browsing agents index a catalogue — but anyone selling llms.txt optimisation is selling you the keywords tag.
• Agent manifests. agents.json, .well-known agent cards, MCP discovery endpoints: real specifications, approximately zero consumer shopping-agent adoption. The actual agentic-commerce rails run through your platform's backend, not your storefront's files. On Shopify your catalogue reaches them through Shopify. Check the toggle in admin and move on.
• hreflang on a single-locale site. It exists to disambiguate variants. One version of everything means nothing to disambiguate.

Measure it, or you are guessing

• Build a query set of 20–50 prompts a real buyer would type, in their language, spanning categories, comparisons and constraints.
• Record who gets cited today, per assistant, before changing anything. Without a baseline you cannot tell improvement from noise.
• Track it separately from organic search. They correlate loosely, and optimising one does not reliably move the other.
• Watch referrals from chatgpt.com, perplexity.ai and claude.ai. Small numbers, but real visits that convert differently.

The uncomfortable part

Everything above is the floor. The ceiling is somewhere else.

Across 75,000 brands, the strongest correlates of AI visibility are all off-site: branded web mentions at 0.664 against 0.218 for backlinks, roughly three times the correlation. And the sources these systems lean on are a recognisable few. Ahrefs measured citation share across 76.7 million AI Overviews, 957,000 ChatGPT prompts and 953,500 Perplexity prompts:

Note that the platforms disagree with each other — Reddit and Quora carry real weight in AI Overviews and do not crack the top ten on either of the others. A source strategy built on one platform's leaderboard misfires on the rest. And these are correlations, and the schema study is the standing warning of what that is worth: cited pages there were nearly three times more likely to carry JSON-LD, and that advantage produced no causal effect at all once it was tested against controls. Large brands are mentioned more and cited more because they are large.

So run the checklist. It is cheap, it is real, and it stops you being invisible on technical grounds. But be clear about what it buys: eligibility, not preference. The store that gets cited is the one people were already recommending to each other, somewhere the models could hear it.

No snazzy claims. Just the checklist.

Sources

A checklist that tells you to distrust unsourced numbers owes you its own.

• Ahrefs, we tracked 1,885 pages adding schema — the controlled study, and the searchVIU crawler test showing five engines extracting visible HTML only.
• Ahrefs, top 10 most-cited domains in AI assistants — the citation-share table, from June 2025 data.
• Ahrefs, AI Overview brand visibility across 75,000 brands — the 0.664 web-mentions figure, and the finding that the top three factors are all off-site.
• Google Merchant Center, how to use conversational attributes — the authoritative list of the six.
• Google Search Central, changes to FAQ and HowTo rich results for the 2023 restriction, and the deprecation note on the FAQPage documentation for the May 2026 removal.

Punde Commerce builds and takes to market custom Shopify stores for premium brands. This checklist comes from production work on a live storefront — every verdict tested against primary sources current as of mid-2026.