A company can appear in an AI answer and still have little say in how it is described. In a September 2026 review of Google AI Mode responses about 25 companies, every company’s website was cited at least once. Yet only 69 of 232 displayed source links, or 29.7%, pointed to the company’s own site. Another 69 pointed to aggregators, directories, retailer pages, and similar profiles.
That does not tell us what fraction of each answer came from those pages. A citation is a visible link, not a measurement of influence inside the model. It does tell us where an audit should begin: with the actual answer and the source trail a buyer can see.
What did 232 citations reveal about company websites?
The 69 own-site links represented 30% of all displayed citations when rounded. Other company-controlled pages, including LinkedIn and parent or investor sites on separate domains, accounted for 47 links (20%). Aggregators and directories accounted for another 69 (30%). News, public-sector, trade, review, and forum sources made up the remaining 47 (20%). Percentages are rounded from the counts.
The own-site share varied from 9% to 58% across companies, with a median of 33%. LinkedIn was the most frequently cited single outside domain, appearing in 20 links. A brand team checking only its website would have missed much of the material displayed beside these answers.
Among the 69 own-site links, 17 went to About pages, five to homepages, and one to a top product or service page. The other 46 led to history, leadership, news, location, and similar pages. Those counts reflect the questions asked. An identity question naturally draws identity pages; it would be a mistake to conclude from this test that product pages rarely matter in product or buying questions.
Where did AI Mode identify the wrong company or an old one?
It described the intended company in 24 of 25 named-company answers. The exception was a community bank with a name shared by unrelated banks. AI Mode returned a list and put a different institution first. At least six of its 11 displayed sources concerned the other banks.
Other answers carried uncertainty into the response. One construction company’s headquarters was hedged between two cities. Another firm’s ownership was described as belonging to its founding family “and potentially” employees. A fiber manufacturer that had adopted a new name in 2025 was identified correctly when named, yet the unbranded industry answer used its former brand and omitted the new one.
Those are different audit findings. The shared name points to entity confusion. The two hedges point to conflicting public accounts. The rename points to a gap between being recognized by name and being discovered under the current name. Each calls for different follow-up.
Can buyers find you when they do not know your name?
Across the five unbranded industry answers, 16 of the 25 studied companies appeared. Furniture surfaced all five; community banking surfaced one. That spread is striking, but there was only one category prompt per industry. Wording and geographic scope differed, so these figures cannot rank the five industries or establish a stable inclusion rate.
The useful comparison is within each company: can AI Mode describe it when prompted by name, and does it include it in a relevant category answer? An accurate branded answer can coexist with absence from discovery. The renamed fiber company illustrated the gap particularly clearly.
What should an AI brand audit actually test?
Start with the questions buyers, candidates, and partners would ask, then save the exact prompts and full answers. Test identity, ownership, capabilities, category membership, and comparisons. Run branded and unbranded versions. Check more than one AI system and repeat prompts so a single response does not masquerade as a pattern.
For each answer, record whether the company is included, where it appears, how it is categorized, which facts are current, what is hedged, and which sources are displayed. Compare the response with a dated, approved account of the organization. Then inspect the cited pages, including third-party profiles and old announcements. A link count alone cannot tell you whether the description is accurate.
This is the Map phase of TILTD’s Interpreter Architecture: identify the human and machine sources shaping how the organization is described. The Brand Interpretation Assessment applies that diagnosis to model responses and the surrounding interpreter network. Model Response Optimization (MRO) then measures interpretation and inclusion over time, so teams can tell whether a change in content or public signals altered the answers that matter.
The immediate output should be a short, prioritized list: wrong entity, wrong fact, outdated name, unclear category, or absence from a relevant discovery answer. Trace each issue to observable sources where possible. Fix the highest-impact contradiction, then rerun the same prompt set. That is more useful than treating a visibility score as the whole diagnosis.
Frequently asked questions
Does a 30% own-site citation share mean 70% of the answer came from elsewhere?
No. It means 69 of 232 displayed source links pointed to companies’ own websites. The study did not measure how much text or reasoning came from each link.
Is an accurate answer by company name enough?
No. This sample included companies that were described correctly when named but were absent or differently named in an unbranded category answer.
How often should a brand rerun an AI audit?
Establish a baseline, then rerun the same important prompts after a rename, major positioning change, source correction, or material model update. Compare answers before drawing a conclusion from a single response.
See what AI is saying before you try to change it
Your website may be in the citations while an old profile, a shared name, or a category answer still changes the story buyers receive. TILTD’s Brand Interpretation Assessment maps the answers, checks where interpretation drifts from your current position, and identifies what to investigate first. Request a Brand Interpretation Assessment at tiltd.ai.