AI Visibility Diagnostic: Why Assistants Overlook Your Brand

Is the brand truly underrepresented?

A brand is underrepresented only when it stays absent from prompts it genuinely qualifies for, tested repeatedly across more than one assistant. One bad answer proves nothing. Neither does a prompt asking about a market you don't serve or a use case your product was never built for.

The reason a single test is worthless comes down to how these systems generate answers. In research by Rand Fishkin of SparkToro and Patrick O'Donnell of Gumshoe.ai, 600 volunteers ran 12 identical prompts a combined 2,961 times, covering ChatGPT and Claude with Google's AI in the same study, and the chance of getting the same brand list twice came in under 1 in 100.

That number reframes what counts as evidence. If a list rarely repeats, your absence from one run is noise, and your presence in one run is equally meaningless as proof of health. What you're measuring is a rate. Set your baseline as the share of runs where you appear, scoped to prompts a real buyer in your market and maturity band would type.

Which prompts should the audit test?

Build the prompt set from how customers describe their problem. People type situations and constraints into an assistant.

The gap between the two languages is measurable. Semrush cross-referenced ChatGPT prompts against a database of 27 billion keywords and found that between 65% and 85% of prompts couldn't be matched to any traditional search keyword at all. The average ChatGPT prompt runs 23 words against roughly 3.4 for a Google search.

So a keyword list ported straight into an AI audit tests a language your buyer doesn't speak. Write prompts as full sentences with the constraint baked in and phrase them neutrally so you aren't leading the model toward your own brand. Run each one the same way across every assistant you care about. Repeat each prompt enough times that you're reading a pattern rather than a coin flip. Ten runs per prompt per assistant is a workable floor for a manual audit.

Do category prompts include the brand?

Unbranded discovery prompts are the honest test, because they ask the assistant to choose without knowing you exist. Tie each one to a use case you actually serve and a buyer constraint you actually meet.

Here the same research is encouraging. Despite wildly different phrasing across 142 human-written prompts about headphones, Bose and Sony still appeared in 55% to 77% of the 994 responses, with Sennheiser and Apple in the same band. Intent survived the chaos of phrasing.

Which means the pool of brands an assistant will consider for a category is far more stable than the order it lists them in. Your goal is membership in that pool. If you're missing from category prompts across 30 or 40 runs, you're outside the consideration set entirely, and no amount of reordering work will help you.

Do comparison prompts exclude the brand?

Comparison prompts tell you whether an assistant treats you as a credible option or as a footnote. Ask for alternatives to a named competitor and ask for a shortlist.

These prompts matter more than their volume suggests. Ahrefs analyzed 26,283 source URLs and found that recently updated comparison listicles were the most prominent page type in ChatGPT's sources, with all of the top 1,000 most-cited pages per platform including them.

The practical read: comparison answers are assembled largely from other people's shortlists. If you appear in category prompts but vanish the moment a competitor is named, the problem sits in third-party comparison content where you're either missing or ranked last. That's a placement problem, and it belongs to whoever owns outreach at your company.

Do branded prompts describe it accurately?

Branded prompts expose whether the assistant knows who you are at all. Ask what the company does and how it differs from the obvious alternative, then read the answer for wrong categories and dead products.

Errors here are common enough to expect them. Metricus, an AI visibility reporting company, reported that 72% of brands it audited had at least one factual error surface in AI responses because of conflicting sources and stale training data.

Sort your errors by type, because the fix differs. A wrong founding date is a source problem you fix in one or two places. A wrong category is an identity problem, and it will keep reappearing until the descriptions on your review profiles and press boilerplate say the same thing. Log the exact wording of each wrong claim. You'll need it later to find which source fed it.

What evidence should teams record?

Record enough per response that someone who wasn't in the room can re-derive your conclusion. For every prompt and assistant, capture whether you were mentioned and where in the answer. Record whether the description was accurate and the tone of it. Log which cited sources the assistant cited and which competitors appeared, along with the date of the test.

The cited sources column is the one most audits skip and the one that pays. AirOps analyzed 21,311 brand mentions across ChatGPT and Claude, with Perplexity included, and found 85% came from third-party sources, with only 13.2% from brand-owned domains.

Given that split, your citation column is a map of who is speaking for you. Save the complete response text, not a summary, because a summary written by the person who already has a theory becomes evidence for that theory. A spreadsheet with response excerpts pasted in survives a budget conversation. A memo saying "we don't show up much" does not.

Which patterns reveal the likely blocker?

Read clusters of misses. Every missed mention looks like a content gap if you squint, which is why most audits end with a content recommendation regardless of what the data said.

The four patterns below each produce a different signature in your evidence table, and each routes to a different team:

• Missing from questions competitors answer, with your own pages uncited anywhere

• Present but miscategorized, with the same wrong description repeating across assistants

• Absent while assistants cite the same reviews and publications featuring rivals

• Strong pages that rank on Google and still never appear as a citation

Ahrefs found that 28% of ChatGPT's most-cited pages have zero organic visibility in Google search, and 65.3% of the cited pages that do rank sit on domains rated 81 or higher. That combination is your diagnostic hinge. Google performance and AI citation are decoupled enough that "we rank well" tells you nothing about which of the four blockers you have.

Missing relevance suggests content gaps

If competitors answer a buyer question and you don't, the shortfall is coverage on your own site. Look for questions where an assistant cites three or four sources, none of them yours, and where you have no page addressing that question in a directly quotable form.

Age matters as much as existence. Seer Interactive analyzed more than 5,000 URLs and found 65% of AI bot hits targeted content published within the past year. The share within two years was 79%, and only 6% landed on content older than six years.

So a page you published in 2021 that still ranks is functionally invisible to the retrieval layer. Before you commission new articles, check publication dates on the pages you already have. Rewriting a stale page that already carries links is cheaper than starting over, and it addresses the freshness signal and the extractability problem in the same pass.

Confused associations signal positioning problems

When assistants put you in the wrong category or describe you for the wrong buyer, the failure is in how the web collectively defines you. Models resolve entities before they retrieve, so a brand that reads as two different things gets described as an average of both.

Wikidata is the lever most teams overlook. It feeds Google's Knowledge Graph, and its notability requirements are lower than Wikipedia's, according to Beatrice Gamba, who teaches entity SEO and knowledge graph strategy at MLforSEO.

The cheap first move is an audit of every place your company describes itself, from G2 to your press boilerplate. If those descriptions disagree about your category, no structured data will fix the conflict. Make them match first, then anchor them with links.

Competitor dominance indicates external weakness

When the same review sites and publications keep surfacing with competitors inside them and you outside, the missing ingredient is third-party validation. More pages on your own domain won't close that gap, because assistants weight what others say about you above what you say about yourself.

Muck Rack's May 2026 analysis of over 25 million links across ChatGPT and Claude, with Gemini included, found earned media drove 84% of all AI citations, a figure that has held between 82% and 89% across three editions since July 2025.

Read your citation column as a target list. The specific domains the assistants keep returning to for your category are the ones worth pitching or reviewing. That's a narrower and more useful brief than "do more PR," and it hands your communications team a ranked set of publications backed by evidence rather than a wish list.

Unseen pages indicate technical barriers

If you have strong, current pages that never surface as citations, check whether the assistants can read them at all. This is the blocker that hides best, because the page looks fine in a browser and performs fine in Google.

The rendering gap explains most of it. A joint Vercel and MERJ analysis of more than 500 million GPTBot fetches found no evidence of JavaScript execution, and the same held for ClaudeBot and PerplexityBot, as well as Meta-ExternalAgent and Bytespider. Googlebot renders. They don't.

That single difference produces the exact symptom you're diagnosing. A client-side rendered page ranks on Google because Googlebot executed the script, and stays invisible to ChatGPT because GPTBot received an empty shell. Fetch your top pages with curl and read the raw HTML. If your product details or pricing only appear after JavaScript runs, no content or PR investment will rescue those pages.

How should the audit score blockers?

Score each blocker on four dimensions from zero to three, and refuse to record a score without a link or a pasted response excerpt beside it. Unsourced scores are opinions wearing a number.

The four columns:

1. Evidence strength: how many prompts and assistants show this pattern

2. Prompt impact: how close the affected prompts sit to a buying decision

3. Competitor disadvantage: how far ahead rivals are on this specific dimension

4. Fix confidence: how sure you are the fix produces a change you can measure

Sample size drives the first column, and the Fishkin and O'Donnell research suggests 60 to 100 runs per prompt before trends stabilize. Most in-house teams won't hit that manually, which is worth admitting in the worksheet rather than hiding.

Cap your evidence score at 2 for anything you tested fewer than 20 times. A blocker scoring 12 out of 12 on thin sampling is a hypothesis with good handwriting. Scoring honestly protects you later, when the fix underperforms and someone asks how confident you were.

Which fix deserves investment first?

Fund the blocker that shows up across the most prompts and the most assistants, and that sits closest to a purchase decision. Breadth of evidence beats severity of any single bad answer, because breadth is what survives the variability you've already accounted for.

Routing follows the pattern. Content gaps go to owned-content repair. Wrong categories go to entity and description cleanup. Competitor-heavy citations go to external authority work. Uncited strong pages go to engineering. Given that only 12% of URLs cited by AI tools overlap with Google's top-10 organic results, per Ahrefs' analysis of 15,000 queries, resist the instinct to route everything to the SEO team by default.

Where evidence conflicts across assistants, fund more testing. Conflicting signals between ChatGPT and Google's AI mean you're looking at two different retrieval systems rather than one broken thing, and picking a side too early is how budgets get spent on the wrong department.

How can Snoika deepen the diagnosis?

Snoika is the step you take when manual testing stops being able to produce the sample sizes your own scoring worksheet demands. Running 60 to 100 repetitions per prompt across four assistants by hand is arithmetic that doesn't work for a marketing team with other obligations.

Snoika launched its SaaS platform in June 2026 with a free AI Visibility Monitoring feature, built for founders and CMOs tracking how AI systems present their brand. The platform covers prompt monitoring and competitor comparison. It also includes citation analysis and sentiment tracking, with prioritized insights on top.

What that changes about your diagnostic is continuity. A manual audit is a photograph of one week. Continuous prompt monitoring turns your four blockers into tracked lines that either move after a fix or don't, which is the only way to find out whether the work you funded actually did anything.

Start with the free monitoring tier and load the prompt set you built from customer language. Compare its baseline against the manual evidence you already collected. If the two agree, you have a blocker worth funding, and now you have a measurement that keeps running after the invoice is paid.