The Practical GEO Audit Checklist: 30 Signals to Test

Access and discovery

1. Canonical response: the preferred URL returns HTTP 200 without a redirect chain.
2. Robots policy: intended search and answer-engine crawlers are not accidentally blocked.
3. Sitemap coverage: every indexable canonical page appears once in an XML sitemap.
4. Internal discovery: valuable pages receive descriptive links from relevant hubs.
5. Document readability: PDFs contain text or have accurate OCR; image-only evidence is not the sole source.

Verification: crawl from an external environment, record the final status and canonical, compare the inventory with the sitemap, and manually inspect representative documents.

Identity and provenance

6. Publisher: the responsible organisation is visible and linked.
7. Author: material content identifies a person or accountable editorial team.
8. Expertise: biographies or review notes explain why contributors are qualified.
9. Method: original research explains sample, collection period, exclusions, and limitations.
10. Corrections: readers can identify material revisions or a correction policy.

Claim quality and sourcing

11. Specific subjects: important statements name the organisation, market, product, or population.
12. Units and periods: figures state what is measured and when.
13. Adjacent sources: a primary source sits beside the claim it supports.
14. Source quality: the link resolves to the original dataset, paper, filing, or official documentation where possible.
15. Claim boundaries: estimates, forecasts, and observed facts are labelled differently.

Page structure and answer clarity

16. Unique title: the title states the subject and user outcome.
17. Descriptive H1: one primary heading matches the page purpose.
18. Question-led sections: subheadings reflect the decisions readers need to make.
19. Answer-sized passages: major questions receive a direct, self-contained response.
20. Useful formats: comparisons use tables; sequences use ordered steps; definitions are explicit.

Metadata and machine-readable consistency

21. Canonical tag: it is self-referential on the preferred URL.
22. Description: it accurately describes the page rather than repeating a generic slogan.
23. Article data: headline, author, publisher, and dates match visible content.
24. Breadcrumb data: hierarchy is both visible and represented accurately.
25. Social metadata: share title, description, URL, and image are complete.

Freshness, monitoring, and evidence of change

26. Date integrity: published and modified dates reflect meaningful content changes.
27. Reference health: material outbound sources still resolve and remain current.
28. Snapshot history: score, page count, failures, and evidence changes are stored by date.
29. Live query set: repeatable prompts cover branded, category, comparison, and evidence-seeking intent.
30. Verification plan: every recommendation has an owner, target state, and retest.

Signals 21–25: AI crawler policy

Access is the precondition for everything else, and it is the check most often skipped because everyone assumes someone else did it. Resolve each AI user-agent individually rather than reading the file and forming an impression.

• Search crawlers reachable: OAI-SearchBot, Claude-SearchBot and PerplexityBot can fetch your pages. These fetch to cite, and blocking them removes you from answers.
• On-demand fetchers reachable: ChatGPT-User, Claude-User and Perplexity-User retrieve a page because someone asked about it right now. Blocking these is the most direct route to invisibility.
• Training policy is deliberate: GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot and the rest reflect a decision you actually made, not a default you inherited.
• No accidental wildcard capture: a broad Disallow in the * group is not silently removing crawlers you intended to allow.
• Policy matches intent: the resolved outcome, checked against the live file, is the one you would describe if asked. Hand-written robots files disagree with their authors more often than expected.

The free crawler access checker resolves all 28 against a live file, and the policy generator composes a corrected block.

Signals 26–30: answerability

The final group is the hardest to automate and the most valuable. These ask whether a retrieved passage can stand on its own once separated from the page around it.

• Standalone opening: the first 200 words answer the page's central question without requiring the heading for context.
• Self-contained claims: material statements survive being lifted out of the page. A sentence beginning “This means that…” does not.
• Adjacent provenance: each significant figure sits in the same passage as its source and the date it was checked.
• Explicit comparisons: where you claim to be different or better, the axis of comparison is stated rather than implied.
• Stated limitations: the page says what it does not cover. Scope statements substantially reduce the risk of an engine misapplying your content to a question you did not answer.

Test the last one by hand. Copy any paragraph containing a number into a blank document and read it cold. If you cannot tell what it refers to, where the figure came from, or when it was true, an answer engine cannot either.

Turning thirty checks into a decision

A checklist that produces thirty equal-weight ticks produces no decision. Weight the results by consequence.

Blocking failures — crawl access, rendering, canonical conflicts — invalidate everything downstream and should be fixed before anything else is measured. There is no value in improving passage structure on a page no crawler can fetch.

Template failures — a missing author field, an absent date, an unlinked methodology — usually affect every page built from that template. One fix, hundreds of pages. These almost always outrank individual page improvements on effort-to-impact.

Page failures are genuinely per-page and should be ranked by the commercial value of the page, not by how many checks it failed. A priority page failing three checks deserves attention before a low-value page failing twenty.

Re-run the full audit after remediation and compare dated snapshots. An audit that is never repeated is an opinion; two dated audits are evidence that something changed.

How to prioritise the audit

Rank issues by business value, affected-page count, expected signal improvement, implementation effort, and confidence. Template-level fixes usually deserve priority because one change can repair many pages. Access failures outrank cosmetic improvements because inaccessible evidence cannot be retrieved at all.

Common questions

How often should a full audit run?+

Quarterly for the full thirty signals on a stable site, monthly if you publish frequently or run a large template estate. What matters more than frequency is that the audit is repeated at all — a single audit is an opinion, two dated audits are evidence of change.

Can this checklist be automated?+

Roughly two thirds of it. Crawl access, rendering, structured data, dates, authorship and link integrity are all mechanically checkable. The answerability signals — whether a passage stands alone, whether a comparison states its axis — need human judgement, and automating them badly produces confident nonsense.

Which signals matter most if I only fix five things?+

Crawler access for the search and on-demand fetchers, a visible and accurate modified date, named authorship, claims adjacent to their sources, and a standalone opening paragraph. Those five clear the most common blocking failures and cost very little.

What if my site fails most of the checklist?+

That is normal on a first audit and it is good news, because early failures are usually template-level. One fix to a shared template typically resolves the same signal across hundreds of pages, so the first audit tends to have the best effort-to-impact ratio you will ever see.

Does a low audit score mean I will not be cited?+

No. The checklist measures whether your content is usable as evidence, not whether it is authoritative. A recognised source with poor structure still gets cited; the checklist tells you how much easier you could be making it, and where you are losing comparisons you should win.

Should I audit PDFs and documents too?+

Yes, and they are frequently the worst offenders. Scanned image-only PDFs contain no extractable text at all, so an engine sees an empty document with a filename. If a PDF carries your most authoritative evidence, it deserves the same audit as an HTML page.