AI does not know you, or AI knows you and does not put you forward: two findings with opposite fixes
Brand recognition and AI visibility are different measurements with opposite fixes. How the branded cohort and the three-way fact check tell them apart.
By Michael Goldrich, Vivander Advisors. Published September 27, 2026.

A resort GM reads an audit that says the property is absent from four of five AI platforms when a guest asks for a pet-friendly resort with a spa. The GM’s first question is the right one: does the AI not know we exist, or does it know and leave us out?
Those are two different findings. They look identical in an unbranded visibility score, and they call for opposite work. The first is a recognition problem: the AI has no stable idea of the property, and the fix is to give it one. In the second, a visibility problem, the AI holds a picture of the property and the picture is missing the facts that would make it a pick. Fixing the first when you have the second wastes a quarter.
The Vivander Visibility Audit separates them with a second cohort of questions that most tools do not run, and with a fact check that most tools cannot run because they never established the facts.
The branded cohort
Every audit carries a branded cohort of at least 20 prompts, built across ten fact dimensions: identity, offer, specialization, market served, geography, scale and status, commercial terms, credentials, proof, and reputation and risk (Observed: Vivander audit methodology, framework v3.14 to v3.65). Each prompt in it names the property. “Does the Larkspur Resort have a spa?” “What is the pet policy at the Larkspur?” “How many people can the Larkspur seat for a private dinner?” “Is the Larkspur a good choice for a corporate retreat?”
The prompts are not invented. Each one is derived from the Verified Fact Block the audit builds at intake, so every verified fact has a question behind it. A fact with no prompt would be a hole nobody asked about. An earlier version of the audit scored only the facts that happened to surface in unbranded answers, and it scored clean on facts nothing ever asked about. The branded cohort closed that gap.
One rule governs the whole cohort. Branded prompts are recorded and never scored. They enter no numerator and no denominator of the visibility score, because a question with the property’s name in it is trivially easy to appear in, and the inflation would scale with the size of the cohort (Observed: audit methodology). Recognition is a reading, never a score.
Reading recognition against visibility
Put the two cohorts side by side and four cases appear.
- Known and put forward. The AI answers the branded questions accurately and puts the property forward on unbranded ones. Recognition and visibility are both present, and the work is maintenance.
- Known and passed over. Branded answers are accurate. Unbranded answers never include the property. The AI holds a picture of the property and does not select it, and the fix is the reason: the fact, the corroboration or the listing that would give the AI a sentence of justification.
- Unknown and absent. Branded answers fail or borrow another property’s facts, and unbranded answers never include the property. Identity comes first: the site, the Google Business Profile and the directories have to agree on what the property is before any of them can say why it is a pick.
- Unknown and put forward. Branded answers fail while unbranded answers include the property. Rare in our audits, and when we have seen it the property was riding on one strong third-party source (Observed: Vivander audits, 2026). Fragile.
What AI knows about your business
The audit’s “What AI Knows About Your Business” artifact runs every verified fact through three comparisons (Observed: audit methodology, framework v3.15).
AI against the verified fact. Does the answer state the fact correctly, incorrectly, or not at all? The judge has an absent state, because “the AI did not say” is a different finding from “the AI said something wrong.” An earlier version without the absent state called true claims wrong when the AI had simply not addressed them, and that version is gone.
Platform against platform. Where two platforms both give an answer and the answers are incompatible, that is recorded as a disagreement. A platform with no answer surface is silent, not dissenting.
AI against the client’s own site. Does the property’s website state the fact? The highest-yield row in the whole audit appears here.
The unsourced-but-true row
A fact the owner confirmed at intake that the website never states anywhere. The spa is real. So is the pet policy. The private room seats 40. None of it is on the site in plain text, and the AI, which can only repeat what it can read, either says nothing or repeats a rival’s figure.
It is the highest-yield row for a reason that has nothing to do with AI. Publishing fixes it, and publishing needs no third party. No directory to petition, no reviewer to persuade, no trade title to pitch. A page, a paragraph, a table of facts, and the fact exists where a crawler can reach it. In our audits, this row is where the Implementation Plan’s first week goes (Observed: Vivander audits, 2026).
Why accuracy is recorded and never scored
A tempting design would score the property on how accurately AI describes it. The audit refuses, and the reason is a trap. Scoring accuracy rewards invisibility. A property the AI never describes has no inaccuracies, so it would score perfectly on accuracy while scoring zero on visibility, and a composite that blends the two would reward the property for staying hidden. Information Accuracy was removed from the Live Visibility Score in version 2.0 for exactly that reason (Observed: audit methodology). Accuracy is recorded in full, row by row, and reported. It moves no score.
A worked example (illustrative)
The Larkspur, a mountain resort three hours from Denver: 140 rooms, a full-service spa, dogs welcome in designated rooms, a private dining room that seats 40. All three of those facts are true. None of the three appears on the website. The spa has a page with a hero photo and the words “renew and restore.” Its pet policy lives in a PDF linked from the reservation confirmation email. The private room appears in a gallery caption. Every figure in this section is illustrative.
Branded cohort, three of the twenty prompts:
| Branded prompt | ChatGPT | Gemini | Perplexity |
|---|---|---|---|
| Does the Larkspur have a spa? | “The Larkspur offers a spa experience,” no details | Correct, citing a regional travel blog | Correct, citing the same blog |
| What is the pet policy at the Larkspur? | “No pet policy information available” | “Pets are not permitted,” which is a rival’s policy | “Contact the resort” |
| Private dining capacity at the Larkspur? | Absent | Absent | “Group dining available,” no number |
Recognition reading: partial. The AI knows the Larkspur is a mountain resort with a spa, learned from a blog. On pets it is silent or wrong. For group dining it has nothing.
Unbranded cohort, one prompt: “pet-friendly resort with a spa within three hours of Denver.” The Larkspur is absent on four of five platforms and mentioned without a reason on the fifth. In the influence ledger, the rivals that were recommended have a pet policy page, a spa page with treatment names and hours, and a groups page with capacities, and the platforms cited those pages directly.
Diagnosis: recognition partial, visibility absent, cause on the site. Three unsourced-but-true rows, three plan items, all within the property’s reach, all publishing.
The fix is three pages: a spa page that names treatments, hours and booking terms; a pet policy page stating the rooms, the fee and the rules; a groups page stating the private room’s capacity, layouts and minimums. Then the same facts go onto the Google Business Profile and the directory listings, so the site’s statement is corroborated. After that, the monitoring window shows whether the unbranded rate moves.
What to do with this
Write down the ten facts about your property that a guest would need in order to choose you. Then open your own website, signed out, and find each one in plain text. Where you cannot, you have found an unsourced-but-true row without an audit.
Ask any AI platform the branded question for each fact, in a fresh session. Where it is silent, the fact is unread. If it is wrong, the fact is being drawn from somewhere else, and that somewhere else is often a listing you can edit.
Then separate the two questions in every conversation about AI. “Does AI know us?” and “Does AI put us forward?” have different answers, different evidence and different fixes. A score that cannot tell them apart will send your team to the wrong work.
Let’s look at how your property is showing up today.
Fifteen minutes. Bring the questions your guests ask. We’ll show you what AI answers.
Book a 15-minute call
