Enterprise SEO Audits, in the Age of AI

How we measured. Google Search Console, property indexableai.com, 90 days from 11 June to 8 September 2026, 1,063 matching query rows. AI citations from Ahrefs Brand Radar across seven engines, joined to 223 Search Console pages on 7 September 2026. A live Ahrefs SERP fetch for “enterprise seo audit”, United States, 9 September 2026, returning 25 rows across the top 12 positions. Search Console positions are impression-weighted and blended across device and country, so treat anything beyond page one as directional. SERP composition on a low-difficulty term is unstable and geo-anchored. One domain, one window, one results page — a case study, not a law.

What is an enterprise SEO audit in the age of AI?

---

An enterprise SEO audit is a diagnostic that answers three questions at scale: can machines reach the estate, can they understand it, and can they reuse it. The first two are familiar. The third is new, and it is where most templates stop.

Retrieval, in this context, means a system fetching a passage from your page to compose an answer — a citation in an AI Overview, a quoted paragraph in ChatGPT, a source link in Perplexity. Retrieval is not indexing. Indexing puts your URL in a database. Retrieval lifts a chunk of your page out of that database and puts it in front of a buyer who may never see your domain name.

The gap becomes measurable at enterprise scale. On a forty-page site the ranking list and the citation list look identical. On a two-hundred-page estate you can count the disagreement, and we did: 17 pages in Google's top 10, 3 of them ever cited (Indexable, 2026).

So the audit acquires a second half. The crawl half asks whether a bot can fetch, render and index the URL. The retrieval half asks whether an engine can find the answer inside the page, isolate it and attribute it. Only one of those is in the standard checklist.

A crawl-clean audit passes pages no engine will ever quote

---

Every one of the 17 pages we held in Google's top 10 was crawlable, indexable and rendering correctly (Indexable, 2026). A conventional audit would have marked all 17 green. Three were cited by an AI engine; 14 were not.

The inverse is harder on the checklist. Our two most-cited pages rank at positions 30.5 and 28.1, which is page three of Google (Indexable, 2026). No crawl-side metric flagged either page as important, and none explains why engines kept choosing them.

Our own tooling makes the blind spot easy to see. The generate_audit_report method in our technical SEO agent returns three KPIs: indexability ratio, crawl health score and AI visibility score. Feed it a worked example — not our estate, but a set of inputs chosen to look healthy — and it returns a crawl health score of 83 out of 100 and an indexability ratio of 90.0% against a 95% target. It also returns an AI visibility score of 100 out of 100, flagged “good”, because no retrieval data was supplied and the parameter falls back to its default.

That default is the argument in one line of code. When retrieval is not measured, the report does not say “unknown”. It says everything is fine. Read any audit deliverable with that question first: which green scores were measured, and which were assumed?

What this results page says about the state of the art

---

The market for enterprise audit advice is thinner than its search volume suggests, and the results page proves it. On 9 September 2026 the US results for “enterprise seo audit” opened with an AI Overview at position 1 carrying three sitelinks, followed at position 2 by screamingfrog.co.uk on a domain rated 87 (Ahrefs, 2026).

Then the authority falls away. Positions 7 to 10 are held by primaryposition.com at domain rating 43, truperformance.us at 26, thenovalab.com at 41 and brandauditors.com at 29, whose ranking pages draw an estimated 46, 55, 29 and 22 monthly organic visits respectively (Ahrefs, 2026). A domain rated 26 whose page attracts 55 visits a month is holding position 8 on a commercial enterprise term.

The demand is real. “Enterprise seo audit” returns 700 US searches a month at keyword difficulty 0 with a traffic potential of 500; its parent topic “seo audit” returns 18,000 searches at difficulty 73 and a traffic potential of 251,000 (Ahrefs, 2026).

Read those together. A term with commercial intent and no competitive barrier is being served, across half the top 10, by pages almost nobody visits — a signal that the enterprise audit question has never been answered properly in public. Difficulty 0 still does not mean rankable: our own page on this term sits at position 27.0 over 90 days (Indexable, 2026). A soft results page says the answer is missing, not that the ranking is free.

Can ChatGPT do an SEO audit?

---

Partly, and the boundary is worth stating precisely, because “Can ChatGPT do an SEO audit?” is a live People Also Ask question on this results page (Ahrefs, 2026). A general-purpose model does one part of the job unusually well and three parts of it not at all.

What it does well is read a single page the way a retrieval system reads it. Paste a page in and a model will tell you whether the answer sits near the top, whether a claim survives being lifted out of its paragraph, and whether the headings match questions a buyer would type. That judgement is hard for a crawler and easy for a language model.

What it cannot do is the enterprise part. It cannot crawl two hundred thousand URLs and hold the graph. It cannot read your server logs, the only place AI crawler traffic is visible. It cannot query Search Console. And it cannot tell you whether an engine cited you, because that means querying the engines and recording what came back.

It also carries a failure mode that matters in an audit: it will produce a confident number that does not exist. Ours did. A drafting pass on our own pricing work invented a competitor's price as a currency conversion (Indexable, 2026). Use a model for the extraction judgement on individual pages, never for the numbers.

What a retrieval audit checks that a crawl audit does not

---

Every crawl-layer check has a retrieval-layer counterpart, and the counterpart is almost never in the template. The left column below is standard. The right column is what has to be added.

Run the left column first, because a page an engine cannot fetch cannot be quoted. To show the shape of that output: hand our calculate_indexability_ratio method an illustrative estate of 219 pages, 208 indexable and 197 indexed, and it returns 90.0% — a five-point gap against the 95% target and 11 pages to fix.

The right column turns the report into a decision. Schema is a retrieval lever, not a ranking one, and a modest one: the measured JSON-LD lift is 6.5 percentage points, with tables and lists adding 2.9 (AirOps, 2026). Implement it as hygiene, then spend the time saved on right-column questions.

How do you prioritise what the audit finds?

---

Prioritise by which layer a finding blocks, not by the tool's severity colour. A crawl-layer error stops retrieval entirely. A retrieval-layer error only stops the citation.

The categorize_issues method in our technical SEO agent sorts findings into errors, warnings and notices and returns a crawl health score out of 100 — ten points off per error, three per warning, one per notice. Hand it one indexing error, two warnings and one notice and it returns 83, a demonstration rather than a measurement. That is a useful triage signal and a poor executive metric: nothing in its inputs describes retrieval, so the score cannot fall when a retrieval problem worsens.

Run two lists instead. The first is the crawl blocker list, ordered by how many URLs each error removes from the index. The top entry on ours was self-inflicted: eleven video pages carrying noindex, applied deliberately because they sent an agency signal we did not want (Indexable, 2026). Deliberate exclusions belong on the list with their reason, or the next auditor will “fix” them.

The second is the disagreement list: every top-10 page with zero citations, ordered by impressions. Ours ran to 15 entries out of 18 (Indexable, 2026). Those pages already hold the ranking; what they lack is extractability. Review it at 45 days, not 30 — citation lag has a median of 6.81 days and a 90th percentile of 37.10 days (Profound, 2026).

Audit your own numbers before you audit your site

---

The most valuable check in our audit process is not aimed at the site. It is aimed at the report. A claim-provenance gate reads every sentence carrying a statistic and asks one question: can we show where this number came from?

Run against our own published pricing work, it caught two defects the structural checks — readability, schema, heading quality — had all passed clean: a competitor's price invented as a currency conversion, and a citation panel that could not be reproduced from any saved report (Indexable, 2026). Structural checks score shape. They have no opinion on whether a number is real.

An independent verification pass then traced every external claim on that page to its primary source. Seven of seven survived exactly, including the labour-cost loading factor, the salary band and the zero-click rate. The remaining defects were arithmetic, not fabrication: a monthly in-house cost figure corrected from about $62,000 to about $65,000 because the annual column included tooling and the monthly column did not, and a three-year cost of ownership corrected from $2.3M to $2.4M (Indexable, 2026, verification record dated 4 September 2026).

That is the honest shape of the result. Most of the sourcing was sound, two numbers were wrong, one correction moved the figure in our own favour, and nothing else in the stack would have caught either.

How do you run the audit?

---

Eight steps, about a week on a mid-sized estate. The first four produce the finding that justifies the rest.

1. Step 1 — start by taking the crawl baseline. Compute the indexability ratio and crawl health score, and list every URL excluded from the index with its reason.
2. Next, build the disagreement list. Export your top pages from Search Console by impressions, pull the URLs AI engines cite, and join them. Every top-10 page with no citation goes on the list.
3. Step 3 — test bot access per agent. Check GPTBot, ClaudeBot and PerplexityBot separately from Googlebot, in robots.txt and at the CDN. A rule written to stop scrapers blocks the engines you want citing you.
4. Verify AI Overview presence live. For your top 20 terms, run a live SERP fetch with a positive control. You should not source presence claims from a cached features field; tool and SERP disagree.
5. Apply the extraction test. Check that ten stat-bearing sentences stand alone — subject, claim and number in one sentence, never a bare opening pronoun.
6. Schedule the freshness sweep. Flag every page outside the 30–89 day update window (AirOps, 2026), then choose refresh or archive.
7. Run the provenance check on your own reporting. Every first-party number needs a record naming the tool, the parameters, the date and the sample size.
8. Implement the re-audit at 45 days, not 30. Citation lag has a 90th percentile of 37.10 days (Profound, 2026), so a 30-day verdict grades work that has not landed.

Fewer than five of these eight means the audit is measuring the crawl and calling it the estate.

In summary

---

An enterprise SEO audit that only tests crawlability is auditing half the system, and the skipped half decides whether a buyer ever sees your name. The next step costs an afternoon: join your top-10 ranking pages to the URLs AI engines actually cite, and count the disagreement. On our estate the answer was 14 out of 17 (Indexable, 2026). Then turn the same scrutiny on the report itself.