Answer engine optimization (AEO): what to do about AI search

A buyer asks an AI assistant for a CRM that suits a small real estate firm. The assistant writes three paragraphs and names two products. Yours is not one of them, and nothing in your analytics shows the question was ever asked.

That is the worry behind answer engine optimization, or AEO: the work of being a source an AI answer draws on or links to. The worry is reasonable. The advice around it is often not, because much of it is stated with more confidence than the evidence supports. This post sticks to what the search and AI companies say in their own documentation, what one large study measured, and a short checklist that follows from both.

Gonzo Digital builds systems that reply to inquiries and connect tools; we do not sell rankings, and nobody can promise a place in an AI answer. We say where we can help and where we cannot at the end.

What AEO is, and what Google says about it

AEO is a label, not a separate discipline with its own rules. Google's documentation on AI features in Search says the SEO best practices still apply to AI Overviews and AI Mode, and that there are no additional requirements and no other special optimizations to appear in them.^[1]

The same page is specific about eligibility. To be shown as a supporting link, a page must be indexed and eligible to appear in Google Search with a snippet, and there are no additional technical requirements.^[1] It adds that meeting the requirements does not guarantee crawling, indexing, or serving.^[1] It also says you do not need to create new machine-readable files, AI text files, or markup, and that no special schema.org structured data is needed.^[1]

Google describes both features as using "query fan-out," meaning they issue several related searches across narrower topics to build a response, which can surface a wider set of supporting links than a single classic search.^[1] Traffic from these features is counted in the Web search type of Search Console's Performance report, with no separate AI report.^[1]

What the click data shows

Pew Research Center analyzed the March 2025 browsing of 900 US adults, covering 68,879 Google searches. In that data, 18% of searches produced an AI summary, and users who saw one clicked a regular result link on 8% of visits, against 15% for users who did not.^[2] They clicked a link inside the summary on 1% of visits.^[2] Searches that started with a question word such as who, what, when, or why produced a summary 60% of the time.^[2]

Read this with its limits. It covers Google only, US adults, and a few weeks in 2025, and AI search products have changed since. It does not say how many people ask AI assistants instead of searching. It does say that, for the searches where a summary appeared, fewer people clicked through, and that being cited did not bring many clicks. Treat visits from AI answers as a small but possibly well-informed share of your traffic, and measure it yourself.

How each AI product finds your pages

Each company publishes the crawler names it uses, and the controls differ by purpose.

• Google uses Googlebot for Search, and its AI features rely on the same crawling and indexing. robots.txt rules for Googlebot are the control for Search, and Google-Extended exists for limiting training and grounding in some of Google's other systems.^[1]
• OpenAI uses OAI-SearchBot to surface sites in ChatGPT's search features, and GPTBot for content that may be used to train its models. The settings are independent, so a site can allow the first and disallow the second.^[3] ChatGPT-User fetches pages for specific user actions and is not used to decide what appears in search; OpenAI says robots.txt rules may not apply to it.^[3]
• Perplexity uses PerplexityBot to surface and link sites in its search results and says it is not used to crawl content for training foundation models. Perplexity-User supports user actions and generally ignores robots.txt rules.^[4]
• Anthropic uses ClaudeBot for training data, Claude-SearchBot to improve search result quality, and Claude-User when a person asks Claude something that needs a page. Anthropic says disabling Claude-SearchBot may reduce visibility in search results, and disabling Claude-User may reduce visibility for user-directed search.^[5]

The practical point is that blocking everything "AI" in robots.txt can also block the bots that decide whether you appear in answers. Decide separately about training and about search, and write down why.

A checklist that follows from the evidence

1. Be open to crawlers and indexed. Check robots.txt and any firewall or hosting rules so the search crawlers above can fetch your pages. Google lists this first among its fundamentals.^[1]
2. Put the answer in text on the page. Important content should be available as text, and structured data should match what is visible, according to Google.^[1] A question as a heading with a direct answer under it is a reasonable layout for readers. The earlier version of this post claimed a 40 to 60 word answer works best. We found no vendor that states a length. For context, the median AI summary in Pew's data was 67 words long.^[2]
3. Say something others do not. Google's guidance on helpful content asks whether your content shows first-hand experience and depth, and whether you are mainly summarizing what others say without adding value.^[6] Original numbers, your own process, and clear limits are what a summary cannot copy.
4. Keep facts current and dated. If an AI answer repeats a price or a rule from your page, it should be right. Put a date on facts that change, and link the source.
5. Keep your business details consistent. Google lists keeping Merchant Center and Business Profile information current among its fundamentals.^[1] The same name, address, and descriptions across your own pages and profiles help people and systems alike.
6. Do not mass-produce pages. Google says using generative AI tools to produce many pages without adding value for users may violate its spam policy on scaled content abuse.^[7]
7. Skip the new files. Google says no AI text files or special schema are needed.^[1] Other products may treat such files differently; we found no primary source showing that they matter.

How to measure it without a special tool

"Share of model" is a name for how often an assistant names you for questions that matter to you. You can estimate it by hand, and no vendor documents a count of it, so treat any tool's figure as that tool's estimate.

List 20 questions a customer asks before buying. Once a month, ask each of them to two or three assistants (our note on Claude vs ChatGPT for business compares two of them) and record whether you were named, linked, or absent, and what the answer said about you. Keep the date and the assistant's name next to each result, because answers vary from run to run. In Search Console, watch impressions and clicks over time, since AI feature traffic is reported inside the Web performance report.^[1] And add one question to your inquiry form: "How did you hear about us?"

Where AI agents help, and where they do not

The old version of this post described agents that watch rankings and rewrite pages on their own to win back a lost "answer position." We found no source showing that this works, and Google's spam policy makes automatic, high-volume rewriting risky. We removed it.

Jobs where automation is a fair fit:

• Checking. A weekly script that lists pages with broken links, missing titles, stale dates, or a robots.txt change. It flags; a person decides.
• Logging. Running your 20 questions on a schedule and storing the answers, so you can see changes over time.
• Collecting. Pulling the questions that arrive in your inbox and form into one list, a good source of topics you should write about.
• Replying. When a visitor does arrive and asks, a reply within minutes matters more than the exact rank. Our note on how fast to reply to a new inquiry covers what the studies do and do not show.

Plain flows in Make, Zapier, or n8n can do the first three; see our guide to no-code tools. Things automation should not do: publish unreviewed text, invent statistics, or edit a live page without a person reading the change.

One speculative note. Some commentators expect customers' assistants to talk to businesses' systems directly to check availability or place orders. We found no source with adoption figures, so treat that as a possibility, not a trend. If you have prices and availability in a format other software can read, you lose nothing by being ready.

What we do and do not offer

We do not promise rankings or a place in AI answers. Our SEO audits read your own site and propose changes that a person approves, and the page says plainly that no citation monitor is connected. Two of our systems relate to this topic. The Lead Response Engine answers the inquiries that search and AI answers send you, and the Connected Operations Layer is meant for letting AI tools read your data through MCP. Compared with an agency retainer, the trade-offs are on our agency retainer comparison.

Questions people ask

What is the difference between SEO and AEO? SEO aims at ranking in a list of results. AEO aims at being a source an AI answer uses or links to. Google says the same fundamentals apply to both.^[1]

Will AI agents replace a marketing team? We found no evidence for it. Automation helps with checking, logging, and collecting, as above, and a person still decides what gets published.

Should I block AI crawlers? That is your decision, and the crawlers do different jobs. Search crawlers decide whether you can appear in answers; training crawlers do something else. Allow or block each on purpose.^[3][4][5]

Do I need an llms.txt file or special schema? Google says no for its AI features.^[1] We found no primary source showing that other products rank pages by such files.