Which AEO Metrics Are Actually Worth Tracking?

Published 2026-09-10 · 6 min read · by the isaeo team

Why this is hard to measure

With SEO you get rankings, impressions and clicks. With AEO you mostly get nothing. If ChatGPT summarises your page in an answer and the user never clicks, that citation is invisible to your analytics — no referrer, no session, no event.

Some answer engines do send referral traffic when a user follows a link, and that's measurable. But it undercounts badly, because the entire value proposition of an answer engine is that the user often doesn't need to click.

So the honest approach splits into two: measure the inputs precisely, because you control them, and treat the outputs as directional rather than exact.

Leading indicators you control

These are countable from a crawl and move when you do work:

1. Structured-data coverage. What share of content pages carry a useful schema type? This is a large component of most AEO scores, and it's binary per page — easy to drive to 100%.
2. Quotable lead passage coverage. What share of pages open with a self-contained paragraph? See writing one.
3. FAQ or question-heading coverage. Explicit FAQ markup where it fits, and question-shaped headings elsewhere.
4. AI crawler access. Binary and critical: are GPTBot, ClaudeBot, PerplexityBot and Google-Extended allowed? Blocking them makes every other metric moot — in this product's scoring, a blocked key bot caps the whole AEO score at 40 regardless of how good the content is.
5. Extractable structure. Share of pages with at least one list or table.

The virtue of these is that they're actionable this week. A composite AEO score built from them — like the one in this product — is a reasonable single number to trend, provided you can explain its parts.

Lagging indicators

Slower, noisier, and closer to money:

• Referral traffic from AI hosts. Segment your analytics by referrer for the known assistant domains. It undercounts, but the trend is informative.
• Branded search volume. If people encounter you inside an AI answer without clicking, some fraction later search your name directly. A rising branded-search line with flat rankings is a plausible AEO signal.
• Direct traffic to deep pages. Unusual, and often means someone was told about a specific page.
• Manual spot checks. Ask the assistants the questions your pages answer, and note whether you're cited. Unscientific, small sample, and still the most direct evidence available.

Run the spot checks on a fixed list of questions each month so the comparison means something.

What not to bother with

1. "AI visibility scores" from tools that won't explain the method. If you can't reproduce the number, you can't act on it.
2. Counting your own crawler hits as AI interest.
3. Obsessing over one assistant's answer to one query. Outputs are non-deterministic; a single miss means little.
4. Waiting for perfect attribution before doing the work. The input metrics are unambiguous and the work is cheap — most of it is editing text you already have.

That last point is the practical one. AEO measurement is genuinely immature, but the fixes are the same ones that make pages clearer for humans, so the downside of acting before the analytics catch up is close to zero.