How AI Answer Engines Pick Products: AEO for D2C Stores

If you run growth or performance for a D2C brand, you have probably noticed a new kind of referral in your analytics: shoppers arriving from ChatGPT, Perplexity, Gemini and Google's AI Mode, often with a very specific product already in mind. The awkward part is that most of the levers you know from SEO and paid search only partly apply. This guide breaks down how answer engines actually choose which products and pages to mention, what the best public data says about it, and the technical checklist we use when we audit a store's AI visibility.

TL;DR for busy operators

• AI answers are built from a different candidate set than Google's blue links. Profound's analysis found only about 12% overlap between ChatGPT's sources and Google's results for the same prompts.

• Classic authority signals transfer poorly. In the same research, traffic explained roughly 5% of citations, and backlinks showed a slightly inverse relationship.

• What does correlate: titles, slugs and snippets that closely match the shopper's intent, fresh content, comparison-style pages, and clean product data that crawlers can read without running JavaScript.

• For product discovery, feeds now matter as much as pages. Google's AI shopping answers are grounded in Merchant Center data, and ChatGPT shopping reads merchant product feeds.

• Measure it like a channel: a fixed prompt panel, share of answer per engine, citation sources, and answer accuracy, tracked weekly.

1. What happens between a prompt and a product mention

It helps to think of every answer engine as a four-stage pipeline: rewrite the prompt, retrieve candidates, select what to cite, then write the answer.

Query rewriting and fan-out

The shopper's prompt is rarely searched as typed. Engines rewrite a long, conversational prompt into several shorter searches. In his BrightonSEO talk, Josh Blyskal of Profound described ChatGPT compressing a multi-sentence prompt into a handful of keywords before it searches. Google describes the same idea for AI Mode as "query fan-out". A prompt like "breathable linen shirt for a Goa wedding, under ₹3,000, not see-through" becomes separate searches for linen shirts, wedding guest outfits, fabric opacity and price bands. Your page competes on each sub-query, not on the original sentence.

Retrieval: each engine looks in a different place

• ChatGPT search retrieves through search partners and OpenAI's own index, built by the OAI-SearchBot crawler. OpenAI states that sites blocking OAI-SearchBot will not be shown in ChatGPT search answers. Profound's experiment of launching a site that was indexed by Google but not Bing, which ChatGPT then could not find, is a good reminder to treat Bing indexing as table stakes.

• Perplexity runs its own crawler and index and, according to Profound, leans heavily on embedding-based matching, so semantic closeness between your content and the query matters a lot.

• Google AI Overviews and AI Mode sit on Google's index, and for shopping they are grounded in the Shopping Graph, which is built largely from Merchant Center feeds.

Selection: the engine only sees a thin slice of your page

At the selection stage, the model often works from very little: a URL, a title and a snippet or description, plus whatever passage it fetched. That is why Profound's advice to "spoil your content" in the meta description works. A description that states the answer gives the selector a reason to pick you.

Ahrefs' study of 1.4 million ChatGPT prompts puts numbers on this. Cited pages had titles with noticeably higher semantic similarity to the prompt (cosine similarity of 0.602 vs 0.484 for pages that were retrieved but not cited), and URLs with natural-language slugs were cited 89.78% of the time when retrieved, versus 81.11% for opaque slugs.

Synthesis: lazy, in a useful way

Profound's line was that "AI search is lazy". When asked for the best running shoe, an engine does not read every brand's product page. It grabs a well-structured comparison that has already done the work and summarises it. In their dataset of 177 million citations, listicles and comparative content made up about 32% of citations, against roughly 9% for standard blog posts.

2. Why your SEO dashboard will mislead you

The most uncomfortable findings from Profound's research are about transfer. Only about 12% of ChatGPT's sources overlapped with Google's results for the same prompts, and only about 26% with Bing's. Site traffic explained roughly 5% of citation likelihood, and sites with fewer backlinks were, on average, cited slightly more often.

Treat these as directional, not laws. They come from one vendor's sampled prompts, engines change their pipelines every few months, and correlation is not causation. But the direction matches what we see in D2C audits: a small, well-structured brand guide can outrank a marketplace in AI answers, and a category leader with excellent SEO can be nearly invisible in ChatGPT.

The practical takeaway is to stop assuming your Google rankings describe your AI visibility. Measure it separately.

3. The D2C layer: pages, feeds and other people's opinions

For product questions, answer engines pull from three kinds of sources, and D2C teams usually only manage the first.

• Your own pages: product pages, collections, buying guides and comparisons.

• Structured product data: Merchant Center feeds for Google, and merchant product feeds for ChatGPT shopping. OpenAI's product feed specification requires fields such as id, title, description, link, image, price and availability, and accepts refreshes as often as every 15 minutes. OpenAI says Shopify merchants' product data is already integrated.

• Third-party opinion: listicles, review sites, YouTube and Reddit. Ahrefs found Reddit URLs are retrieved at scale but rarely cited directly. They make up 67.8% of non-cited URLs in its sample. That suggests Reddit shapes what the model believes far more than it shows up as a link.

One structural shift matters here. In March 2026, OpenAI pulled back from processing purchases inside ChatGPT and refocused on product discovery, with merchants keeping their own checkout. For D2C brands, the battleground is getting into the shortlist that the answer presents. The click and the sale still happen on your store.

4. The technical checklist

Let the right bots in, deliberately

Search crawling and model training are separate decisions. Allowing OAI-SearchBot lets you appear in ChatGPT search answers; GPTBot is OpenAI's training crawler; and ChatGPT-User fetches pages when a user asks ChatGPT to, which OpenAI says may not follow robots.txt. A sensible default for most D2C stores looks like this:

Check that your CDN or bot-protection layer is not silently blocking these user agents. That happens more often than robots.txt mistakes.

Assume the crawler does not run JavaScript

Most AI crawlers read the HTML they are served and do not execute client-side JavaScript. Shopify's Liquid themes render core product data on the server, but many apps inject reviews, size charts, FAQs and delivery estimates in the browser. If those are where your differentiators live, the answer engine never sees them. Test it the way a crawler would:

If your rating, review count or size guide does not show up in the raw HTML, move it server-side or into the page's JSON-LD.

Get indexed by Bing, on purpose

Verify the store in Bing Webmaster Tools, submit your sitemaps and consider IndexNow for faster recrawls of price and stock changes. It is a 20-minute job that many D2C teams have never done.

Make product data machine-readable and consistent

Every product page should carry Product structured data with an Offer (price, currency, availability), brand, identifiers such as GTIN where you have them, and aggregateRating backed by real reviews. Add MerchantReturnPolicy and shipping details, because "easy returns" and "delivery by Friday" are exactly the constraints shoppers put in prompts. Keep page, structured data and feed in sync. Google is increasingly strict about mismatches between Merchant Center and the live site.

Write titles, slugs and descriptions for the selector

• Use descriptive product handles: /products/white-linen-shirt-regular-fit, not /products/sku-4471-wht. Redirect old URLs when you change them.

• Front-load the attributes shoppers ask about into titles: material, fit, use case and occasion.

• Make meta descriptions answer the question: "Breathable, opaque 100% linen shirt in a relaxed fit, made for humid weddings. Free size exchange." beats a brand slogan.

If you want to be rigorous, embed your top 50 shopper prompts and your product titles with any open embedding model, compute cosine similarity, and rewrite the titles that score lowest against the prompts they should win. That is a crude proxy for what the selector sees, and it is a much better starting point than gut feel.

On llms.txt

The talk was bullish on llms.txt, a markdown map of your key pages placed at the root of your domain. It is cheap to ship, and harmless, but evidence that major engines rely on it is thin, and Google has publicly played it down. Ship it if it takes an hour; do not expect it to move results on its own.

5. Content that answer engines actually cite

For D2C brands, the content plan follows from the "lazy engine" finding:

• Comparison and "best for" pages: "Best linen shirts for humid weather", "Cotton vs linen for office wear", "Brand X vs Brand Y". Be honest, include alternatives, and put the comparison in plain HTML lists rather than images or scripts.

• Buying guides that answer the fan-out: one section each for occasion, fabric, fit, budget and care, mirroring how engines split the prompt.

• Freshness as a habit: Profound saw a page drop from first to fifth after competitors refreshed theirs, and recover within a week of being updated. Ahrefs' median cited page was around 500 days old, so you do not need daily churn. But review your top guides quarterly and update facts, prices and the year.

• Show up where opinions form: in Profound's cloud-computing example, a brand that joined the handful of Reddit threads shaping its category saw its Google AI Overviews visibility rise from 23.8% to 73% within a month. Participate transparently, as the brand, and add real information, not spam.

6. Measure it like a channel

You cannot optimise what you do not track, and AI answers are volatile. A minimum viable setup:

• A fixed prompt panel: 50 to 100 real shopper prompts across categories, occasions and price bands, run weekly on ChatGPT, Gemini, Perplexity and Google AI Mode.

• Share of answer: the percentage of prompts where your brand or product is mentioned, broken down by engine and by competitor.

• Citation sources: which URLs, yours and third parties', each engine cites for each prompt. This tells you which listicle or thread to influence next.

• Answer accuracy: wrong prices, discontinued products and outdated return policies are common, and they cost conversions.

• Logs and analytics: watch server logs for OAI-SearchBot, ChatGPT-User and PerplexityBot hits, and create an AI referrals channel in GA4. ChatGPT tags many outbound links with utm_source=chatgpt.com.

7. A 30-day plan

• Week 1: build the prompt panel and baseline your share of answer per engine. Check robots.txt, CDN bot rules and Bing indexing.

• Week 2: run the curl test on your top 50 products, move critical attributes server-side, and fix structured data and feed mismatches.

• Week 3: rewrite the lowest-scoring titles, handles and meta descriptions. Publish two comparison or "best for" guides for your biggest categories.

• Week 4: engage in the top third-party threads and lists shaping your category, refresh your strongest guides, and re-run the panel to measure the change.

Where Eldor fits

This is the work our AI visibility audit automates for D2C stores: it runs real shopper prompts across the major answer engines, shows where your products are and are not being recommended, which sources the engines cite instead, and which product data is missing or inconsistent. If you want a baseline before you start, run a free audit.

Sources: Josh Blyskal, Profound, "We Analyzed 10,000,000 AI Prompts" (BrightonSEO, 2025); Ahrefs, "Why ChatGPT Cites One Page Over Another"; OpenAI crawler and commerce documentation; Search Engine Land on ChatGPT Instant Checkout. Figures are from those studies' samples and will shift as engines evolve.