98.6% of visible citations in a recent OpenAI local-business study came from third-party websites.
That is the sort of number that can change an AEO strategy overnight.
It can also become misleading very quickly if the measurement contract is not visible.
The study was sent to me by James Tandy, founder of Empirank, after he read our AI Search Visibility Measurement Guide. Empirank analysed 2,403 visible citations across 300 local-business answers from what it describes as a search-enabled OpenAI configuration.
Only 1.4% of those citations were attributed to the recommended businesses’ own websites.
The immediate marketing conclusion seems obvious:
A business cannot depend on its website alone. It also needs accurate listings, reviews, local portals, professional associations, publishers and other independent sources.
I think that direction is reasonable.
But I would not place the 98.6% figure beside another AI citation percentage until I had compared the measurement designs.
What exactly does 98.6% measure?
The denominator is 2,403 visible citations.
It is not 300 businesses.
It is not 300 recommendations.
It is not the percentage of answers influenced by third-party sources.
It is the share of recorded citation instances classified as third-party.
That distinction matters.
The 300 answers contained an average of about eight citations each. But averages can hide an uneven distribution.
Imagine that 30 long answers produced most of the citations while the other 270 answers displayed only one or two sources. The pooled citation percentage would then describe the citation-heavy answers more strongly than the typical answer.
A more complete report would show at least two views.
The first is citation-level composition:
Of all recorded citation instances, what percentage pointed to owned and third-party sources?
The second is answer-level coverage:
What percentage of answers contained at least one owned source, at least one third-party source, both, or neither?
Both are valid measurements. They answer different questions.
Search-enabled must be observable
There is another issue that has become increasingly important in my own work.
A provider accepting a search parameter does not prove that search occurred.
An API can return HTTP 200 while silently ignoring an unsupported field. A system can also return citations without exposing enough metadata to prove whether they came from live retrieval, a prebuilt index or another grounding layer.
This creates a dangerous experimental failure.
The collector records:
retrieval = on
The provider returns:
status = 200
The analyst assumes the answer came from live search.
But unless retrieval leaves independently observable evidence, “retrieval on” may only be a label created by the experimenter.
A credible retrieval-on study therefore needs to record how search execution was verified.
Depending on the platform, that evidence might include tool-call records, retrieved-source annotations, grounding metadata, provider documentation tied to the exact endpoint, or a controlled freshness test.
HTTP success is request evidence. It is not retrieval evidence.
Ownership also needs a contract
“Owned website” sounds like a simple classification until real businesses enter the dataset.
A company may operate through:
its main corporate domain;
a country-specific domain;
a verified subdomain;
a parent-group website;
a local franchise website;
an authorized booking platform;
a marketplace storefront;
a social profile;
or a separately operated location page.
Which of these count as owned?
Empirank states that its business-source attribution was deliberately conservative and that exact root-domain matching may miss franchise relationships, verified subdomains and other owned properties.
That is a useful limitation to publish.
It also means the 98.6% result depends partly on the ownership rule.
A conservative matcher may classify some controlled or authorized properties as third-party. A permissive matcher may make the opposite mistake and treat partner-controlled pages as brand-owned.
Neither rule is automatically wrong. The classification just needs to be declared and versioned.
Citation does not mean recommendation influence
The study makes another important distinction: a visible citation does not prove that the cited page caused the recommendation.
That deserves more attention.
A source can be cited for an address, opening time or service description after the model has already selected a business.
Another source may have shaped the shortlist but receive no visible attribution.
A citation can therefore play several different roles:
It may introduce the business.
It may verify a factual claim.
It may support a comparison.
It may provide background information.
It may only confirm a minor detail.
Counting all of these together produces a citation total, but not a causal model of recommendation.
From outside the platform, we can usually measure visible mentions and visible citations.
Source influence is harder.
If a page is cited, we know it was exposed in the answer layer. We do not automatically know that it caused the brand to be selected.
If a page is not cited, we do not know that it had no influence.
Why this cannot be compared directly with China retrieval-off data
Some of my historical work examined Chinese AI API surfaces with retrieval disabled or unproven.
Those observations sometimes produced a very different visible-source pattern.
That is not enough to say one dataset is right and the other is wrong.
The studies differ across several dimensions:
market;
language;
engine;
product surface;
retrieval state;
buyer intent;
business category;
citation extraction;
ownership classification;
and collection period.
If those dimensions are not matched, a percentage difference cannot safely be labelled “engine behavior.”
It could be a retrieval difference.
It could be a prompt-intent difference.
It could be a market difference.
It could be an extractor difference.
It could be all four.
The correct use of the new result is as a configuration-specific external comparison, not as a universal benchmark.
What marketers can act on today
We do not need to wait for perfect attribution before improving the source layer.
The owned website still needs to act as the current source of truth.
It should clearly state the business identity, location, services, categories, official routes, policies and other facts that buyers need to verify.
Third-party sources should then corroborate those facts.
That may include relevant directories, industry associations, review platforms, local publishers, professional profiles and trusted community discussions.
The goal is not to appear on every directory.
It is to build a consistent source portfolio around the decisions that matter to the buyer.
A third-party source that repeats an outdated address or incorrect service category is not an AEO asset just because an AI system can retrieve it.
The right operating question is:
Can a buyer verify the same important fact across the official source and the independent sources most likely to appear during the decision?
That is more useful than chasing a global citation percentage.
A minimum comparison contract
Before comparing two AI citation studies, I would want the following fields:
Collection date.
Market and language.
Engine, model and product surface.
Retrieval state and retrieval-verification method.
Prompt categories and buyer intent.
Number of independent answers.
Number of citation records.
Citation occurrence and deduplication rules.
Domain normalization and ownership rules.
Source-category taxonomy.
Extractor version.
Invalid-answer policy.
Answer-level citation distribution.
Study limitations.
Without those fields, two dashboards may display “citation share” while measuring different objects.
The result may look precise and still be incomparable.
The useful conclusion
The Empirank study adds useful evidence that search-enabled local-business answers can rely heavily on visible third-party sources.
It does not prove that third-party sources caused the recommendations.
It does not establish a universal OpenAI citation mix.
It should not be blended directly with retrieval-off China observations.
Its practical lesson is still valuable:
The website is the official fact layer, but it may not be the only visible verification layer.
Its methodological lesson is even more valuable:
Before comparing AI visibility numbers, compare the instrument that produced them.
Study:
https://empirank.com/studies/openai-local-business-citation-sources/
Visibility Atlas measurement guide:
https://visibilityatlas.com/guides/ai-visibility-measurement/
