What are the biggest AEO mistakes to avoid?

An accidental block can start with a reasonable decision: you don't want your content used to train AI models. The mistake is carrying out that decision with a rule that also covers search.

OpenAI's crawler documentation shows why the two have to be handled separately. It describes OAI-SearchBot as the crawler that surfaces sites in ChatGPT's search features, and GPTBot as the one that collects content for training. Each setting is independent. A site can allow the first and disallow the second. But a site that opts out of OAI-SearchBot won't be shown in ChatGPT search answers, though it can still appear as a navigational link. So a robots.txt rule that disallows both crawlers, written with only training in mind, gives up the search answers as well.

Google has its own version of this. Since August 31, 2026, every Search Console property has a generative AI control that includes or excludes the site from AI Overviews, AI Mode and the generative features in Discover. Include is the default. A property with a parent inherits the parent's choice by default, so a setting made on the parent applies to the child until the child is configured separately.

This exclusion is easy to miss. Google says the control isn't a ranking or inclusion signal for the rest of Search, so a site can be excluded from AI answers and remain eligible everywhere else. The control also doesn't govern training, and Google points to a separate one, Google-Extended, for that. Excluding a site over a training concern gives up the AI features without addressing the concern.

Getting access right removes one barrier, and that's all it does. Google says plainly that indexing and serving aren't guaranteed even when a page meets every requirement. Other products document their own crawlers, and we cover those in which AI platforms an AEO strategy should target.

The second mistake is trying to cover every way a question can be phrased with a separate page. It feels like diligence, and it produces a lot of pages quickly.

Google's guide to optimizing for AI search names this directly. Creating separate content for every variation of a search, done mainly to manipulate rankings or AI responses, violates its scaled content abuse policy. That policy describes many pages generated mainly to manipulate rankings, with little value for the reader, no matter how they were created. So writing with AI isn't the offense, and neither is a second page on a related question. Purpose and value decide it.

The guide also explains why the variants don't help. Google's systems understand synonyms and general meaning, so they can connect a search to a page that doesn't use the same words. Rewording one answer across a dozen pages isn't needed for those searches to find it. What Google says does help is what it calls non-commodity content: a point of view or first-hand experience that the other sources lack.

Suppose your sales team keeps hearing one question: how long will it take to move our 400-page site to a new CMS? Suppose, too, that you have records from past migrations. The formulaic plan writes one page for a 100-page site, another for a 200-page site and one more for each CMS, each restating general advice. The other plan writes one answer from those records. It gives the range of timelines in them, says what stretched the long ones, and sets out an estimate for a 400-page site. That page carries information the general advice doesn't have.

Where the questions come from matters too. Lean Labs' description of its AEO work starts with a managed question set built from sales calls, CRM insights and Search Console demand. It then describes expert-led answers connected to supporting proof. Starting from questions buyers have asked gives each planned page a reason to exist that you can point to.

Structured data is worth having. The mistake is believing it earns citations, and scheduling it ahead of work that matters more.

Google's guide says structured data isn't required for generative AI search and that there's no special schema.org markup to add. It still recommends structured data as part of ordinary SEO, because it helps pages qualify for rich results, and it asks that the markup match the visible text. The same guide says Google Search ignores llms.txt files, and that you don't need to break content into small chunks or rewrite it in a special style for AI. Those statements are about Google Search. Other services can use such files, and the guide says so.

So schema belongs on the list as one item of technical upkeep. The technical step in Lean Labs' process lists schema and entity clarity alongside crawl health and indexing checks. Markup describes a page. It doesn't make an unindexed page eligible, and Google doesn't present it as a route into AI answers. We go further into what markup does in what role structured data plays in AEO.

Someone types a question into an AI product, sees the company named, and takes a screenshot. That's a sample of one answer on one day. Google notes that AI Overviews and AI Mode may use different models and techniques, so the two can show different responses and links. A screenshot from one doesn't tell you what the other showed.

Google's generative AI performance report in Search Console counts impressions, meaning the number of times a link to your site was shown in AI Overviews or AI Mode. You can group them by page, country, date or device. An impression is a link on a screen. It isn't a visit, and it isn't a lead.

The fix is to keep three things in separate columns and report them together: whether you were shown or cited, whether anyone came, and what those visitors did. Lean Labs describes its measurement in those layers: mentions, citations and prompt coverage first, then AI referral traffic and engaged visits, then the contacts, deals and pipeline recorded in the HubSpot CRM. The value of that arrangement is that a rise in the first layer can't pass for a rise in the third.

It also tells you what to investigate when the numbers disagree. If citations rise and visits don't, people could be getting what they need from the answer, the tracking could be missing visits, or the cited pages could be ones a buyer has no need to open. If visits rise and contacts don't, look at who is arriving and where they land, and allow for the time your buyers take to make contact. We describe the tracking setup in how to track traffic and referrals from AI answer engines.