Why a relevant page is not enough: finding where a brand disappears

The sentence “the brand is not visible in AI” does not show where the chain broke. A page may answer the prompt precisely while its domain is absent from the recorded source list. The domain may appear in that list while the short answer omits the brand. The brand may be named without an exact link or recommendation.

Record three events separately: whether the page fits the prompt, whether the owned domain appears in the study’s source record, and whether the answer names the brand. The study’s source record is not necessarily the same as links displayed in a consumer interface. Exact links, visits, and purchases require their own measurements.

• The page fits the prompt.
• The domain appears in the source record.
• The answer names the brand.
• An exact link and a visit are confirmed separately.

Section sources:[1] Benjamin Tannenbaum / arXiv

Benjamin Tannenbaum examined 34,960 repeated GPT and Gemini observations across 2,854 unbranded prompts and 75 anonymised Aiso projects. The observations were collected from June through September 2026 and covered GPT-5 nano, GPT-5.4, and Gemini 2.5 Flash. The starting prompts did not name the target brands.

The paper compared three signals: how closely a public page matched a prompt, whether the owned domain appeared in the recorded source list, and whether the system generated an additional brand-specific search query. It then compared those signals with the final brand mention. This is an observational comparison, not a proven causal chain.

Section sources:[1] Benjamin Tannenbaum / arXiv

The large panel had four states: neither signal; the owned domain alone; the system’s additional brand-specific search query alone; and both signals. For GPT, the brand appeared in 438 of 15,524, 866 of 1,769, 38 of 59, and 117 of 128 answers. For Gemini, it appeared in 524 of 13,801, 1,995 of 3,415, 28 of 33, and 231 of 231 answers.

Those counts equal 2.8%, 49.0%, 64.4%, and 91.4% for GPT, and 3.8%, 58.4%, 84.8%, and 100.0% for Gemini. They are rates inside one proprietary sample, not the probability that a new page will earn a mention. They do not promise higher search positions, links, or sales.

Section sources:[1] Benjamin Tannenbaum / arXiv

In a separate check, the author compared 199 prompts with 275 public pages from one organisation. Page-to-prompt match separated cases with and without the owned domain in the source record more clearly for Gemini than GPT. On the study’s 0.5-to-1 scale, the results were 0.641 and 0.545: 0.5 means no useful separation and 1 means perfect separation.

Useful content therefore matters, but page fit alone did not reliably predict whether the domain appeared. When a relevant page is absent from the source record, the narrow conclusion is that the checked page did not appear in the available record for that run. The cause remains unknown; it cannot be read directly from the final answer.

Section sources:[1] Benjamin Tannenbaum / arXiv

In the 199-prompt subset, GPT recorded the organisation’s domain 30 times and named the brand in all 30. Among the other 169 cases, it named the brand 16 times. Gemini recorded the domain 65 times and named the brand in 48 of them; among 134 cases without the domain, it named the brand 11 times.

This is a strong association within the subset, but it is not a proven mechanism. The domain may have appeared alongside an already selected brand rather than causing the selection. The paper therefore does not support a promise that adding a link will, by itself, produce a brand mention.

Section sources:[1] Benjamin Tannenbaum / arXiv

The subset shows another path: the brand was sometimes named even when its own domain was absent from the source record. At least three explanations remain possible: a third-party page carried the relevant information, the model relied on information learned earlier, or another unobserved path was available. The study cannot choose between them.

For an audit, owned and third-party pages must therefore be marked separately. The absence of a brand-owned link does not prove that the answer had no support at all. A mention without a link also does not prove a recommendation, a visit, or a sale.

Section sources:[1] Benjamin Tannenbaum / arXiv

Create one row for every run. Record the exact prompt, language, country when known, service, model, mode, date, account state, and repeat number. Save the full answer and links. Then mark separately: page fit, owned domain in the source record, third-party page, brand mention, recommendation, exact link, and confirmed visit.

Do not replace unknown with zero. “No link displayed” is an interface observation, not proof that the page was never found. “No visit recorded” means the visit was not observed, not that influence was impossible. For every percentage, show both the matching count and the total checked count.

Section sources:[1] Benjamin Tannenbaum / arXiv

If the page does not answer the prompt well, start with the reader’s task and the content. If the page fits but its domain is absent from the source record, inspect server-rendered HTML (the page code returned immediately by the server), indexability, the preferred page address, internal links, access to key facts, and independent public support. These are hypotheses to test, not a proven ranking formula.

If the domain is recorded but the brand is absent, check whether the page clearly connects the brand, fact, and conditions. If the brand appears without its own domain, look for third-party support and do not credit the result only to the owned site. The paper does not release its underlying data, countries, or account types; its author leads Aiso and reports a financial conflict. The guide can locate the missing checkpoint, but it cannot promise indexing, a link, a visit, or a sale. Telegram is only the short distribution version, not the factual source.

Section sources:[1] Benjamin Tannenbaum / arXiv