What are the biggest AEO mistakes to avoid?

AEO programs underperform for reasons that sit upstream of writing quality. Access, eligibility, and measurement failures decide whether a page can be cited at all, and none of them are visible in the content itself. Google states the eligibility rule plainly: to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, meeting the Search technical requirements [1]. A page that fails that test cannot be rescued by better headings or more FAQ schema.

The access layer got more complicated

Each platform now runs several crawlers with different jobs, and one robots.txt line covers all of them if the user agent is wrong. OpenAI separates OAI-SearchBot, which surfaces sites in ChatGPT's search features, from GPTBot, which collects training data [7]. Anthropic splits ClaudeBot for training from Claude-SearchBot for search indexing and Claude-User for live user-triggered fetches [8]. Google's Google-Extended token governs Gemini Apps training and grounding plus Vertex AI grounding, and does not affect Search inclusion or ranking at all [3].

Content and judgment failures follow

Google's scaled content abuse policy targets pages generated in volume to manipulate rankings rather than help users, and it names generative AI tools as one route to that outcome [2]. Separately, AI referral traffic came to 0.2% of visits across 99 billion sessions in Contentsquare's 2026 benchmark [14], which makes raw volume a poor read on whether the work paid off. The mistakes below are ordered by cost relative to how often they go unnoticed.

AI referral volume is too small to justify a program on its own and too noisy to judge one by. Across 99 billion sessions on 6,500 sites, Contentsquare's 2026 benchmark put AI-referred traffic at 0.2% of total visits, growing 632% year over year and converting at 1.3%, below email at 1.9% [14]. A program measured on session counts against that baseline reads as a failure by month three and gets defunded before it can compound.

The value shows up in composition rather than count. Ahrefs published its own numbers in June 2025: 0.5% of its traffic came from AI search, and that 0.5% produced 12.1% of signups, a conversion rate it calculated as 23 times higher than traditional organic for its business [13]. The two datasets do not contradict each other. Contentsquare reports a cross-industry average, while Ahrefs reports a product whose buyers ask AI assistants exactly the comparison questions its content answers.

The real cost of the volume mindset is misallocation. Teams optimizing for clicks gravitate toward queries that still produce clicks, which pushes effort into the informational middle of the funnel where AI answers now absorb the click anyway, and away from the decision-stage questions where a citation influences a purchase.

The correction is to define the outcome before the tactics. Choose the 30 to 100 prompts a buyer would plausibly type before selecting a vendor in your category, measure how often you appear in those answers, and track pipeline from AI-sourced sessions separately. Session volume becomes a diagnostic input rather than the scoreboard.

Unintentional blocks happen because one rule covers crawlers that do completely different jobs. OpenAI documents the split directly: sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, while disallowing GPTBot indicates only that content should not be used to train foundation models [7]. Anthropic reports the same effect for its own bots, warning that disabling Claude-SearchBot may reduce a site's visibility and accuracy in user search results [8].

The CDN layer reaches further than most teams assume

Cloudflare's July 2026 update sorts AI traffic into Search, Agent, and Training categories and applies the most restrictive matching rule, so multi-purpose crawlers including Googlebot, Applebot, and BingBot are blocked for customers who selected a Training block [9]. From September 15, 2026, new domains get Training and Agent blocked by default on pages that display ads [9]. A setting someone enabled in 2024 to keep content out of model training can quietly cover the crawler feeding ordinary search results.

There is now a switch inside Search Console

Google's search generative AI control lets an owner exclude a site from AI Overviews, AI Mode, and generative features in Discover, after which links to the site will not appear in those features [4]. It does not affect the rest of Search, so a site can vanish from AI answers while its organic reporting looks unchanged.

The correction is a quarterly access audit. Read robots.txt against each platform's published user agent list, check the bot rules in the CDN or WAF, since those usually belong to a different team than the one running content, and use server logs to confirm that search-purpose crawlers are receiving 200 responses instead of 403s.

Publishing hundreds of near-identical question pages puts a site inside Google's definition of scaled content abuse, which covers generating many pages for the primary purpose of manipulating search rankings and not helping users [2]. The policy lists using generative AI tools to produce many pages without adding value as an example, and applies the same standard however the text was created [2].

Authorship is the part teams get backwards. The exposure is not that a system detects machine writing and rejects it. Google's policy turns on whether pages add value, so a well-edited AI-assisted page carrying original information sits outside it while a hand-written page restating public knowledge at scale sits inside it.

Redundancy is expensive in a specific way

An answer engine retrieves a small set of passages per query, so a page competes on whether it holds a fact, number, or first-hand observation the alternatives lack. Three hundred templated pages restating what is already documented elsewhere give a retrieval system three hundred equally skippable candidates. Volume does not raise the odds of selection when every candidate carries the same information.

The cost compounds at the site level. Vercel's crawl data shows AI crawlers already wasting roughly a third of their fetches on pages that return 404 [11], and adding thin pages spends more of that budget on content nobody will retrieve.

The correction is to trade volume for information gain. Publish when a page carries something original: your own data, a documented process with real numbers, pricing you will stand behind, or a limitation nobody else writes down. Fewer pages that each answer one question completely outperform a library answering forty questions partially.

AI answers draw heavily on sources you do not own, so a website-only program competes for a minority of the citation slots. Semrush analyzed more than 100 million AI citations from 230,000 prompts across ChatGPT search, Google AI Mode, and Perplexity in weekly snapshots between July 14 and October 12, 2025, and found Reddit and Wikipedia holding the top two positions among ChatGPT's most-cited domains [15]. Community platforms and reference sites sit ahead of nearly every brand-owned domain in that ranking.

The practical consequence is that a category's buying questions often get answered from a forum thread, a review platform, an encyclopedia entry, or a roundup on a publisher site. If a product is absent from those places, or described with outdated pricing and features, the model's answer inherits the gap regardless of how well the company's own pages are structured.

That pool also moves on its own. Semrush recorded a mid-September 2025 change that dropped Reddit from roughly 60% of ChatGPT responses to around 10%, while Google AI Mode and Perplexity stayed consistent [15].

The correction is to treat off-site work as part of the same program rather than public relations overflow. Identify which third-party URLs are actually cited for your priority prompts, then work those surfaces. Outdated facts on review and comparison sites are usually the fastest fix, since one correction can change how several answers describe you. Participating honestly in the communities where your category gets discussed is slower and much harder to fake, which is part of why it holds up. A model reconciling conflicting descriptions across sources tends to reach for the version that appears most often.

Eligibility breaks. Google's documentation states that a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements, before it can appear as a supporting link in AI Overviews or AI Mode [1]. An engagement that delivers schema, answer-first formatting, and an llms.txt file while leaving a page uncrawlable has optimized something that cannot be selected.

Rendering is the failure that hides best

Vercel and MERJ analyzed 569 million GPTBot fetches and 370 million Claude fetches and found no evidence that any major AI crawler executes JavaScript, even though ChatGPT fetched JavaScript files on 11.50% of requests and Claude on 23.84% [11]. Content that exists only after client-side hydration is absent from what those crawlers read. The same analysis found ChatGPT spending 34.82% of its fetches on 404 pages against Googlebot's 8.22%, so stale links and unmanaged redirects waste far more of an AI crawler's budget [11].

Weakening rank correlation argues for fundamentals, not against them

Ahrefs' March 2026 analysis of 863,000 keyword SERPs and 4 million AI Overview URLs found 38% of cited pages also ranked in the top 10, down from roughly 76% in July 2025, with the rest split almost evenly between positions 11 to 100 at 31.2% and beyond the top 100 at 31.0% [12]. Being indexed and reasonably visible still puts a page in the pool, and a top-three position no longer guarantees the citation.

The correction is sequencing. Confirm that main content renders server-side and that pages are indexed with clean status codes before spending on answer formatting. The technical work is cheaper, it is verifiable, and it gates everything downstream.

Without a baseline there is no way to tell whether AI visibility changed or whether the reporting did. Search Console's generative AI performance report covers impressions from AI Overviews and AI Mode, but does not include click data or average position, and it excludes experiments in Search Labs [5]. Impressions confirm that a link appeared and say nothing about whether anyone acted on it.

Analytics has a matching gap

Google Analytics 4 now includes an AI Assistants default channel covering arrivals from sources such as ChatGPT, Gemini, DeepSeek, Copilot, and Grok, matched on an ai-assistant medium or a recognized AI referrer [6]. That channel explicitly excludes Google's AI Overviews and AI Mode, which arrive as ordinary organic search [6]. Sites that never configured custom channel groupings spent the preceding period filing a growing share of high-intent sessions under direct traffic, which is why so many programs cannot produce a credible before-and-after.

The baseline has to exist before the first change

Record where you stand on a fixed prompt set, capture current generative AI impressions in Search Console, and make sure AI-sourced sessions are separated in analytics with conversions attached. Re-run the identical prompt set on a schedule, since changing the wording alone moves results enough to invalidate the comparison.

Ahrefs' reporting shows why tying sessions to outcomes matters more than counting them: its AI search traffic was 0.5% of sessions and 12.1% of signups, a ratio visible only once AI sessions are separated from direct traffic and joined to conversion data [13].

Citations are not durable, which makes a page that earned one in March an unreliable asset by June. Parse measured 693,509 answers across 16,143 ChatGPT prompts and 15,805 Google AI Overviews prompts between March 26 and April 25, 2026, and found two repeat ChatGPT answers to the same prompt shared only 21.2% of cited domains on average, with AI Overviews steadier at 31.5% [16]. Source churn exceeded 75% in every industry the study examined [16].

Selection rules move independently of your content

Ahrefs' comparison of its July 2025 and March 2026 datasets shows Google pulling far fewer AI Overview citations straight from the original SERP, with top-10 overlap falling from about 76% to 38% [12]. Semrush caught a single mid-September 2025 change that dropped Reddit from roughly 60% of ChatGPT responses to around 10% [15]. Neither shift had anything to do with the content on the affected pages.

Two things follow. Measuring visibility once produces a number with a wide error bar, so a single check showing a brand cited or absent should not drive a decision. And a page that stops being cited has usually not degraded; either the surrounding pool changed or a competitor published something more specific.

The correction is a maintenance cadence rather than a campaign. Re-check the prompt set at a consistent interval, read appearance rate as a trend line rather than a status, and refresh pages carrying dated facts, pricing, or product details before those numbers go stale. Pages built on original data need a review date attached at publication.

The evidence base is uneven by design

Platform documentation is reliable on access and eligibility and nearly silent on selection. Google publishes the bar for appearing in AI Overviews and AI Mode [1] without publishing how a page gets chosen from the eligible pool. Everything about selection comes from third-party measurement of observed outputs, which shifts whenever the platforms change.

Published conversion multiples do not transfer

The 23x figure comes from Ahrefs' own traffic, for a product whose buyers ask AI assistants comparison questions [13]. Contentsquare's cross-industry benchmark shows AI-referred traffic converting at 1.3%, below email [14]. Using any published multiple as a forecast produces a business case that will not survive contact with your own numbers.

Blocking crawlers is sometimes defensible

Not every block is a mistake to reverse. Cloudflare reported OpenAI crawling at roughly 1,700 fetches per referral and Anthropic at about 73,000 to 1 in June 2025, against roughly 14 Google crawls per referral [10]. A publisher whose revenue depends on page views may reasonably conclude the exchange is poor. The mistake is making that call by accident, or applying it to search crawlers when the intent was to limit training.

Vendor measurement is not standardized

AEO tools sample different prompt sets, at different frequencies, from different locations, and report incompatible numbers for the same brand. Given the churn Parse measured [16], low sampling frequency can manufacture trends that are not there. Keeping one tool and one prompt set long enough to build a comparable series matters more than picking the best-scoring tool.

Using AI to help write is not one of them

Google's spam policies target scaled, low-value page generation and apply the same standard however the content was produced [2]. A page carrying original information and edited by someone who knows the subject is outside that policy regardless of what drafted the first version.

A month-over-month drop in AI sessions is not automatically a failure

Repeat answers to the same prompt shared only 21.2% of cited domains in Parse's measurement [16], so short-term movement in AI referrals is often noise. Judge the program against a fixed prompt set across a quarter instead.

Declining model training is not the same as blocking search

The user agents are documented separately by OpenAI [7], Anthropic [8], and Google [3]. A site can opt out of training data collection and remain fully available to the crawlers that feed AI answers, provided the rules name the right bots.

Schema markup is not the missing piece by itself

Structured data helps machines parse a page and does not create eligibility. A page still has to be indexed and snippet-eligible to appear as a supporting link in AI Overviews or AI Mode [1].

Zero citations is not always a content problem

Check access and rendering before rewriting anything. No major AI crawler executed JavaScript in Vercel and MERJ's analysis of nearly a billion fetches [11], so a client-rendered page can be genuinely good and still invisible.