How an answer engine audit actually runs

Updated August 14, 2026

An audit runs in five passes over one crawl: discovery, edge probing, per-page checks, scoring, and a ranked fix plan. It uses no headless browser and makes no model calls at audit time, so it is deterministic — run it twice on an unchanged site and you get the same numbers.

Pass 1 — how does it find your pages?

Sitemap-first. We fetch /robots.txt, follow every declared sitemap (including nested index sitemaps), and fall back to link discovery from the homepage where a sitemap is missing or short. A 679-page production site completed in about ten minutes and roughly five thousand HTTP requests, with polite concurrency and a delay you can raise for fragile origins.

Pass 2 — what does the multi-user-agent probe do?

Your homepage and a sample of content URLs are fetched nine times over, once per real AI crawler user-agent string, plus a Chrome baseline. Each response is graded on three axes:

• Status — anything other than 200 for a bot that a browser gets 200 for.
• Challenge markers — interstitial fingerprints that a non-JavaScript client can never solve.
• Body parity — word count against the browser baseline. A 200 that returns a stub is still a block.

Then the finding is cross-joined against what /robots.txt declares. Agreement is a pass. Disagreement — policy says allow, edge says no — is the highest-value finding the tool produces.

We always caption a bot-level block as detected from a spoofed-user-agent probe rather than from an operator IP. Fake-bot rules can produce the same signature, which is precisely why log verification exists as the tier that upgrades probable to proven.

Pass 3 — what runs on each page?

Every fetched page runs through the check catalog: 61 checks in the current engine build, each a pure function of the raw HTML and response headers. No check depends on a rendered DOM, a screenshot or a model opinion, which is what makes the result reproducible and cheap enough to run on ten thousand pages.

Pass 4 — how is the 0-100 built?

Each check carries a weight inside its category; category scores roll up into the weighted total described on the home page. Alongside the AEO score we compute a separate classic SEO baseline, because search indexability is a hard prerequisite rather than a competing metric: over 87% of ChatGPT citations match Bing's top organic results, Claude aligns about 86.7% with Brave, and Google's AI surfaces require Google indexing. We also derive four per-engine readiness sub-scores from the same crawl at no extra cost.

Pass 5 — what happens to the findings?

Failures are grouped into a fix plan ranked by weight times affected pages, so the first item is always the one that buys the most score for the least work, with every affected URL named. On a fix engagement those become staged edits: each target file is backed up, the diff is generated from your site's own entity data rather than a generic template, and nothing is written to your origin until you approve it. Then we re-run the audit and show you the delta.

Vague suggestions are the single most common complaint against this category of tool. A fix that a developer cannot paste is not a fix; it is homework.

— from our teardown of the five leading paid AEO platforms, August 2026

What do you actually receive?

1. A dashboard with five views: overview, per-page table, ranked fix plan, methodology, and the run log.
2. The raw JSON report — every check result for every page, yours to keep and diff against later runs.
3. The probe matrix: nine user-agents by sampled URLs, with status, challenge and parity for each cell.
4. An off-site action plan covering the factors we deliberately do not score.

Related reading: the full check catalog, what a policy-versus-edge mismatch looks like, and what an engagement costs.