Anonymized first-party research
A GEO content score needs a citation cross-check
A page can score well against a content checklist without earning more AI citations. In our exploratory audit, an appealing external-reference correlation weakened when we checked a different citation surface. The useful result is a measurement lesson, not a new ranking recipe.
Put the citation measure beside the content score
A content heatmap describes features of a page. It cannot, by itself, tell you whether an AI answer used that page. The useful comparison puts the factor scores beside an observed citation outcome, its denominator, and the source record.
Our broader screen scored 3,519 eligible pages across 13 site inventories, then added a two-page supplement. Thirteen account Sources exports yielded 186 eligible cited-page matches. Those are different quantities: we did not obtain 3,521 independent citation outcomes.
The 71 rounded-zero rows were still present in captured citation tables. Dropping them would remove observed cited pages from the analysis. At least ten also had literal own-page links in exported GPT answer text. That is why our private working table keeps three separate states: cited, absent from the captured table, and citation data unavailable.
The pattern that looked promising
One anonymous within-site subset contained already cited blog pages. A higher automated external-reference score tracked higher combined Sources Used percentages: Spearman's rank correlation was 0.51. After adjusting for page length, it remained approximately 0.51.
That looked like a useful hypothesis. It was not a test of whether adding links caused more citations, and it did not compare cited pages with a clean, contemporaneous set of uncited pages. The subset was examined because an exploratory association appeared, so it was not an independent validation sample.
Same blog subset. Different citation measures.
The cross-check changed the story
We separately counted literal own-page URLs in that account's nonempty exported GPT answers. On the same blog subset, the external-reference correlation was 0.05; adjusting for page length left it close to zero.
These measures are not interchangeable. Combined Sources summarizes the app's citation evidence. The answer-text export is a distinct surface, without timestamps sufficient to align every historical observation. The difference does not establish an engine effect, and we cannot attribute it to rounding alone.
It does stop a stronger claim. The apparent relationship did not repeat on this cross-check, so it would be premature to sell adding external references as a proven way to gain citations. Exported Gemini answer text omitted URLs entirely; we treated its text-level citation information as missing, not as zero citations.
A content factor needs its own audit
The external-reference proxy counted descriptive links to other hosts. Inspection found substantive references, but also navigational links, generic resource blocks and a malformed URL. Some short source names were missed. That is not the same as measuring evidence quality.
We tested an automated reconstruction of ten public content criteria. Some original weights and semantic scoring anchors were unavailable, so the version used disclosed assumptions and lexical heuristics. It was not the original author's exact framework or an independently rated quality benchmark.
| Factor | Automated check | Limit |
|---|---|---|
| Answer-first sections | Does the opening sentence overlap the section question? | Lexical overlap is not a human judgment of whether the answer is useful. |
| Question headings | How many headings look like questions? | A question-shaped heading can still introduce irrelevant content. |
| Top summary | Is there an explicit or heuristic summary near the top? | A short opening paragraph is not necessarily a good summary. |
| Self-contained passages | Do paragraphs avoid context-dependent openings? | A pronoun heuristic cannot establish passage independence. |
| Short sentences | How many extracted sentences stay under a word threshold? | Brevity does not establish completeness or accuracy. |
| Lists and tables | Are list items or tables present? | A format can be present without helping the buyer. |
| Statistics and specifics | Do paragraphs include recognized numeric detail? | A number can be irrelevant, outdated or unsupported. |
| External references | Are descriptive links to other hosts present? | A link count is not a validated measure of source credibility. |
| Named expertise | Is an author schema or byline detected? | Detection does not verify identity, credentials or subject expertise. |
| Freshness | How recent is a publisher-declared modification date? | A date change does not prove a substantive content update. |
What belongs in a public heatmap?
The internal research table joins page scores to exact citation URLs. This public version reports grouped findings instead. We have removed company and client names, domains, page titles, URL paths, prompts, locations and distinctive commercial examples. We have also withheld the per-page score matrix: replacing a URL with "URL 1" would not remove the risk of matching a distinctive row back to its source.
No client dataset, raw export, screenshot, hidden page-level array or downloadable evidence bundle is attached to this article. The original records remain private. The chart above describes a measured relationship, rather than exposing the underlying pages.
What this changes in a content plan
Use a score to choose questions to investigate. Prefer evidence that lets a buyer verify a specific claim. A higher checklist score is not a reason to promise a ranking gain.
- Choose comparable pages and relevant prompts before looking at the outcomes.
- Preserve the pre-edit content and citation baseline.
- Change one evidence feature, rather than several formats at once.
- Observe a fixed future window with stable prompts and separate model results.
- Keep citation presence, frequency, brand mentions and customer outcomes separate.
This connects with our analysis of what a website gives AI to verify and our discussion of onsite and offsite evidence. This article's narrower contribution is showing why a within-site pattern still needs a second citation measure and an audit of what the factor counts.
Methods and limitations
Capture dates: 17–18 October 2026. Source types: public page extractions and first-party Aiso Sources workbooks. Eligibility required at least 100 extracted words of HTML content, with utility pages, duplicate canonicals and unavailable bodies excluded. Unknown factors remained missing. The two-page supplement was not a complete additional site inventory.
All ten factors and a composite were screened. Twenty-five primary account/factor comparisons had enough sample and outcome variation to compute under the protocol; a Benjamini–Hochberg adjustment covered that primary family. Later page-type and answer-text checks remained exploratory. A separate implementation independently recalculated the central correlations; there was no independent human semantic rater.
Seven exports retained the app's You classification filter; six included all domain types. Citation absence in a filtered export is not a complete historical zero. Current page content was captured after much of the citation observation period. Canonical matching can merge URL variants, while rounded percentages lose detail.
Length and coarse page-type controls do not remove query relevance, authority, templates, language, publication history or correlated pages. The pooled rounded-zero statistic is page-weighted, is dominated by one account, and is not an average account rate. None of these observations establishes that changing a content factor causes a citation gain.