Skip to content

Case study

I pointed my own GEO audit tool at my own site

Generative-engine optimisation is the question of whether an AI assistant cites you when someone asks it for what you do. I built a tool that measures it, then ran it against my own site on 2026-08-23 for $1.00 of provider spend. Every number below comes from that run. I have not regenerated it to look better.

Cited in 3 of 18 observations. Site score 46.

What was actually queried

Two of the four surfaces were never queried in this run. There was no Perplexity key, and the run predates the OpenAI key. A zero on a surface nobody asked is an absence of evidence, not evidence of absence, so it is reported as not queried rather than as a zero.

  • Google AI Overviews · queried
  • Claude · queried
  • Perplexity · not queried
  • ChatGPT · not queried

Page scores

/
32F
/work
60C
/about
50D
/services
52D

Eligibility

All 44 of 44 technical eligibility checks passed. Each one still carries the severity it would have had if it had failed, which is a field, not a finding. Rendering it as a finding would turn a clean sweep into two dozen invented criticals.

Where the score comes from

Six rubric factors per page, scored out of 100. Three cells were never measured and say so. They are not zeros, and reading them as zeros would invent three failures the tool never found.

Rubric factor scores out of 100 for each audited page. Cells marked not scored were not measured in this run and are not zeros.
PageAnswer StructureCitable SubstanceEntity SchemaAuthority FreshnessExtractabilityTechnical Hygiene
/0062.5505080
/work0100100050100
/aboutnot scored0100not scored50100
/services70083.3not scored50100

Ranked fixes

23 fixes, ranked by impact against effort. The top eight:

  1. 1Statistics density/servicesimpact 75 · effort 3
  2. 2Mainline extraction quality/impact 60 · effort 3
  3. 3Mainline extraction quality/aboutimpact 60 · effort 3
  4. 4Quotable sentences/servicesimpact 50 · effort 2
  5. 5Concrete specifics/servicesimpact 50 · effort 2
  6. 6Original content mass/impact 50 · effort 3
  7. 7Original content mass/aboutimpact 50 · effort 3
  8. 8Original content mass/servicesimpact 50 · effort 3

What the tool got wrong

Publishing the run without this section would make it a demo, not an audit.

  • A coverage false negative. The tool reported zero coverage for prompts it had in fact observed, because an empty result set and an unqueried surface were collapsed into the same shape. That is precisely the confusion this page is careful about everywhere else.
  • A misattributed citation basis. A result row labelled its basis as Google when the observation came from a different surface. A citation attributed to the wrong engine is worse than a missing one, because it survives review looking correct.
  • This run was scored under rubric 1.0.0. The installed rubric has since moved on, and the run is deliberately not re-rendered under it, because rescoring an old artifact under a new rubric produces a number that never existed.

Method and limits

One run on 2026-08-23: 18 observations across 2 of four surfaces, 4 pages, rubric 1.0.0, $1.00 of provider spend. A single run is a snapshot, not a trend. 2 of the 18 calls failed outright and are counted as failures rather than quietly dropped.