When a page answers a buyer's question directly, yet the AI response omits the brand, the default impulse is to rewrite the copy. That impulse skips diagnosis. An absent mention confirms a breakdown occurred somewhere between the prompt and the output, but it does not locate the fault.
Three distinct failure modes require separation:
- Query fit: Does the page actually satisfy the specific constraints of the prompt?
- Observable exposure: Does evidence from or about the business appear in the system's recorded sources?
- Selection and synthesis: Does the model choose to name the brand and represent it accurately in the final output?
Teams routinely inspect the first stage via content audits and track the third via mention monitors. The middle stage often goes unexamined. Treating either end of the pipeline as a complete diagnostic leaves the operational gap between them unexplained.
What observable source records actually measure
An analysis of stage-level AI visibility by Benjamin Tannenbaum examined unbranded prompts across operational GPT and Gemini monitoring logs. The study tracked two intermediate signals: whether the business's own domain appeared in the stored source URLs, and whether an observed search query expansion explicitly included the brand name.
When neither signal appeared in the logs, brand mention rates sat at 2.8% for GPT and 3.8% for Gemini. When the business's own domain appeared in the stored sources without a branded query expansion, mention rates rose to 49.0% in GPT and 58.4% in Gemini.
These figures require strict interpretation. They measure contemporaneous associations within the same run, not the causal lift you should expect from forcing a page into the index. Furthermore, the observable record is partial. A model might extract brand facts from a third-party directory, rely on pre-trained weights, or execute unlogged retrieval steps. The absence of a first-party URL in the stored source record does not prove the system encountered zero evidence about the business.
Methodological note: Tannenbaum leads Aiso Boost, a vendor of AI-search measurement software. Because the underlying client records are proprietary and withheld, the study's statistical models cannot be independently refitted.
Anchor the diagnosis to one commercial prompt
Broad visibility scores obscure specific failure modes. Instead of auditing in the aggregate, select one commercially vital question and log its exact parameters: the prompt text, the full generated answer, the exposed source list, the timestamp, and the model configuration. Collect multiple runs for the same prompt to distinguish stable patterns from stochastic noise.
Next, place the target page next to the exact wording of the prompt. General topical relevance is rarely enough. If a prospect asks for an inventory tool with native NetSuite integration and UK VAT compliance, a high-level feature overview that omits those specific constraints fails at the query-fit stage.
If the page lacks the required specifics, you have identified a concrete content gap. If the page already supplies the exact answer, halt the rewrite and inspect the source record.
Be precise about what your monitoring tool actually captures. If an interface only surfaces end-of-answer links, label them as final citations—do not mistake them for a comprehensive trace of every document the retrieval system evaluated.
Triage by the available evidence
Once query fit is confirmed, the relationship between the source record and the final text points to one of three distinct problems:
- The target page is missing from the observed sources. Focus investigation on crawl access, indexation, and how closely the document structure matches likely query expansions. Treat these as hypotheses to test, not proven causes. Simultaneously, check whether third-party domains in the source list mention the brand.
- The target page is present in the sources, but the brand is omitted from the answer. Being retrieved is not the same as being recommended. Inspect which specific passage the model likely extracted and compare your page against the competing URLs in the source list. The system may have pulled a generic definition or baseline statistic from your page while relying on rival sources to build its shortlist of vendors.
- The brand is named, but the claims are inaccurate. Compare the model's falsehoods against both your page copy and the third-party URLs in the source list. This identifies a representation issue—a failure mode that binary mention trackers completely miss.
A single domain can exhibit different behaviors across prompts and model families. Document that variance rather than flattening it into a single site-wide score.
Associations are diagnostics, not repair recipes
The strongest predictive model in the Tannenbaum paper combined historical mention baselines with live retrieval signals from the current run. Because live source signals only exist after a query executes, the model serves as a post-hoc diagnostic for monitored prompts, not a crystal ball for unmonitored queries.
Similarly, a strong historical mention rate does not isolate why a brand wins. Prior visibility can stem from training-data prominence, widespread third-party citations, or consistent first-party retrieval. Observational logs cannot cleanly untangle those mechanisms.
This reality should constrain what an AI audit promises. Low observed source exposure justifies looking into technical access and query alignment, but it does not guarantee that injecting schema markup, tweaking H2 tags, or publishing supplementary articles will flip the output.
The practical value of this framework is restraint. Rather than commissioning blanket rewrites whenever a brand goes unmentioned, teams can document a precise chain of evidence: the exact prompt, the intended landing page, the observed source log, and the point where the observation becomes uncertain. Any proposed edit should name which link in that chain it targets and how subsequent runs will verify whether the intervention worked.
