Skip to content
All articles
Answer Engine OptimizationAI MarketingSEOE-CommerceReyo

Answer Engine Optimization: A Guide for Retailers

Answer engine optimization for retail: how AI answers get built, which crawlers decide your visibility, what works, and which tactics you can skip.

By Marcel Hühn · · 11 min read

EnglishDeutsch
AI-generated 3D illustration: a shop shelf as a plinth, above it a floating speech bubble holding three brand signs with one of them lit, beside it a block with a magnifying glass and a bookmark

A shopper no longer types "running shoes wide feet" into a search box. She asks an assistant: "Where do I buy running shoes for wide feet that arrive tomorrow?" The answer names three retailers, adds a line about fit and one about the return window. If your shop is not in there, you did not lose a click. There was never a list you could have dropped out of.

Answer engine optimization is the work of being in that answer. This guide covers how such an answer gets assembled, which access you have to open, which levers the vendors themselves describe as working, and which widely sold tactics you can skip. Every statement about crawlers, controls and reports here was checked against vendor documentation on 20 November 2026.

The short answer

Answer engine optimization means making sure an answer engine can find your content, is allowed to use it, and names you as a source. That is less new than the label suggests. In its guide to the generative features of Search, Google states that optimizing for generative AI search is optimizing for the search experience, and thus still SEO, and that AI Overviews and AI Mode carry no additional requirements and need no special optimizations.

Three things do shift, though:

  • The answer is assembled, not listed. There is no position three to settle into any more. There is one paragraph with two to five sources.
  • Several systems, each with its own rules. Every major vendor runs its own crawlers, and you decide access per purpose, not per vendor.
  • Measurement changes. Visibility inside an answer often produces no visit at all, so counting sessions alone hides the channel.

Answer engine, AEO, GEO: the terms

An answer engine answers a question instead of listing results. That includes assistants like ChatGPT and Claude, AI Overviews and AI Mode in Google Search, Perplexity, and Copilot in Bing.

Answer engine optimization (AEO) is the work of showing up as a cited source in those answers. Generative engine optimization (GEO) means the same thing and is the more common label in English-speaking markets. When you compare vendors, do not let the acronym impress you.

Two things are not meant here: creating content with AI, which is a separate question, and measuring your own visibility. Measurement is the report; AEO is the work.

How an answer gets built

In practice three steps run: the engine gathers material, selects from it, and writes an answer with citations. Only two gates matter for you, and both are technical.

  1. Access. May the vendor's crawler read your page?
  2. Permission. May that content be used as a snippet?

Google is unambiguous about its own features: a page qualifies as a supporting link in an AI Overview or in AI Mode if it is indexed and eligible to appear in Search with a snippet. With the assistants, the search crawler decides, not the training crawler. That distinction is the one most often got wrong.

The crawlers that decide

The columns are the real decision. Make it per purpose, and make it deliberately.

Vendor Search and answers Training User-initiated fetch Worth knowing
OpenAI OAI-SearchBot GPTBot ChatGPT-User Block OAI-SearchBot and you are not shown in ChatGPT search answers; the three settings are independent, and a robots.txt change takes about a day to apply
Google Googlebot separate control Googlebot Eligibility is indexing plus snippet permission; nosnippet, max-snippet and noindex limit how content is featured in the AI experiences
Perplexity PerplexityBot not used for foundation models Perplexity-User The user-initiated fetcher generally ignores robots.txt; a blocked page may still surface as domain, headline and a brief factual summary
Anthropic Claude-SearchBot ClaudeBot Claude-User robots.txt is honored, including Crawl-delay; since 30 October 2026 a machine-readable training preference is honored as well, and a rule naming ClaudeBot takes precedence
Bing and Copilot bingbot separate control bingbot NOARCHIVE keeps content out of Copilot responses and grounding, NOCACHE limits use to URL, title and snippet

How to read it: for visibility in answers you need column one and column three open. Column two, training, can be closed independently without costing you citations. Blanket-blocking everything with "AI" in its name locks you out of the answers and buys you a kind of control you never wanted in the first place.

The levers the vendors name themselves

Google's guide is the only detailed first-party source on this question. Translated into shop work:

  • Technical basics. The page is indexed, allowed to show a snippet, and publicly crawlable. Content that only exists behind a "show more" interaction is not text to a machine.
  • First-hand content over summaries. Google explicitly contrasts a review of something you tested yourself with a summary of what is already available. Good news for retail: fit notes, care instructions and range comparisons cannot be copied from anywhere.
  • Images and video. Pages that support their text with relevant, high-quality images and video can gain visibility beyond plain text links.
  • Structured data that matches. It is not required for the AI features, but it still helps with rich results. The condition is that everything in your markup is also visible on the page.
  • Product and business data. Maintained Merchant Center feeds and a current business profile help products and local stores appear in AI responses.
  • Page experience. Rendering across devices, low latency, and a main content area clearly distinguishable from everything else.

For a product range the conclusion is blunt. What an assistant needs to recommend you is not marketing copy but specifications: availability, delivery time, return window, size range, material, price validity. If those live inside an image or under a different label on every page, your shop drops out of the answer even with the better offer.

What you can skip

Google's guide also lists tactics you can leave alone: AI-specific files such as llms.txt, breaking content into tiny pieces for machines, and collecting inauthentic mentions across the web.

The llms.txt case deserves a closer look, because it shows up as a deliverable in a lot of proposals. Google states the file will not positively or negatively impact your visibility or rankings. A Google Search team member publicly called it speculative for now. An Ahrefs analysis of 137,000 sites found that 97 percent of existing files received zero requests in May 2026. We are pragmatic about it at Reyo: our site writes such a file at build time because it costs us one step. We do not count it as a lever, and we would not move budget for it.

Chopping content into tiny answer blocks is overdone too: clear structure with descriptive subheadings helps readers and machines alike, fifty disconnected sentences help nobody.

Access or control: the one real trade-off

There is exactly one place where you have a genuine decision to make. Google's robots meta tag documentation says nosnippet applies to AI Overviews and AI Mode and also prevents the content from being used as direct input for them. On the Bing side, NOARCHIVE keeps content out of Copilot responses. Both work, and both cost you the citations.

So decide per content type instead of across the board:

  • Open: product pages, category copy, buying advice, shipping and returns pages. That is the material you get recommended with.
  • Partly limited: sections written only for signed-in customers. Use the attribute for a single page section, not the whole page.
  • Closed: internal price lists, trade terms, promotions before launch. Those do not belong in a public index anyway.

Measuring without fooling yourself

Four sources, in order of how much weight they carry:

  1. Search Console. The generative AI performance report shows how your content performs in the generative AI features of Google Search and in Discover.
  2. Bing Webmaster Tools. Since February 2026 a public preview report covers AI performance: citations in Copilot and in AI summaries, the query phrases behind them, and counts per URL. Bing notes these are sampled figures that do not indicate ranking.
  3. Your own sample. Twenty questions your customers actually ask, always in the same wording, once a month, and you note which brands get named. That is a sample, not a measurement of the market. It is enough to see a direction.
  4. Referrals in your analytics. They arrive late and in small numbers. Useless as an early indicator, fine as confirmation.

What you cannot measure is your share of all answers. Nobody can, and anyone promising it is selling a sample as a census.

The first four weeks

Week Task Result
1 Check robots.txt against the table above, allow the search crawlers, write down the training decision A file that reflects a decision instead of a copy
2 Write down the twenty questions your customers ask and run the first sample A baseline to compare against in January
3 Audit ten product and category pages for hard specifications: availability, delivery time, returns, sizes, material Facts as text, not as images
4 Set up Search Console and Bing Webmaster Tools, open both AI reports Two reports instead of assumptions

After that it is routine: the sample once a month, the specifications on your most important pages once a quarter.

Common mistakes

  • Blocking every crawler with "AI" in the name. You block the search crawlers too, and you disappear from the answers, not just from training sets.
  • Selling llms.txt as a measure. As things stand, there is no documented effect.
  • Watching clicks only. An answer without a visit is still a recommendation to somebody who is about to buy.
  • Leaving hard facts inside images. Delivery time and return windows as graphics do not exist for an answer engine.
  • Checking once and ticking it off. Vendors change crawlers and controls every quarter, which is why the table above carries a date.

How Reyo helps

For us, AI answers are one observed channel alongside email, push, ads and website promotions. Our AI visibility measurement asks assistants the questions your customers ask and shows, per prompt category, how often your brand and how often your competitors appear in the answer. The feature is in beta, and it is a repeated sample, not a census of the market.

The rest of the market comes in through competitor monitoring, and analytics turns it into series across weeks and months. Two limits belong with that: we do not name the shops, because the point is the patterns rather than the senders. And we see what was sent, not how it performed.

If you would rather not run your sample by hand, the route goes through your own AI client. Reyo MCP is the interface for that: your client, for example Claude, ChatGPT or Claude Code, reads your competitors' campaigns and emails, creates entries in the campaign planner and writes comments, always inside your own account only. Access is included from the Professional plan, and the guide walks through the steps per client. For what such a server actually is, read What is an MCP server?. For rolling the wider topic out in a team without starting at the deep end, read how to use AI in marketing.

Conclusion

Answer engine optimization is not a second discipline next to SEO. It is a shift of attention, from placement to being quotable. The work itself is unglamorous. Search crawlers get access, training crawlers get whatever you decide. The hard specifications of your range sit on the page as text, not inside an image. And what you count is a monthly sample plus two reports, not a share of answers that nobody can actually know.

If you do one thing tomorrow, open your shop's robots.txt. Chances are good there is a line in there keeping you out of answers you could already be winning.

Last checked 20 November 2026. Statements about crawlers, controls and reports come from the documentation of Google, OpenAI, Perplexity, Anthropic and Bing, verified on 20 November 2026.

Reading about the market is good. Seeing your competitors live is better.

30 minutes, your competitors set up before the call, one recommendation to take away.