An AI visibility report should show the losing prompts

An AI visibility report can look reassuring while the business is still absent from the buyer questions that start deals. A headline score rose from 42 to 51. The dashboard is greener. Then sales asks whether the company appears when someone asks for a tool that solves the exact problem it sells, and nobody can answer.

That is not a reporting problem caused by a missing chart. It starts when the report treats every prompt, assistant and outcome as interchangeable. A mention in a broad educational answer and a recommendation in a high-intent comparison are both observations. They are not the same observation.

A useful AI visibility report makes the loss visible enough for a team to investigate. It shows the question, the answer, the competitor that appeared, the evidence a human can inspect and the next action. It also keeps uncertainty on the page instead of hiding it inside an average.

Its job is to connect a buyer question with an observed answer. It is not a replacement for web analytics, a search ranking report or a sentiment survey. Those can sit beside it. They answer different questions.

Start with a decision the reader of the report needs to make. A product marketer may need to choose which comparison page to research. A content lead may need to decide whether a missing source is an access problem or a thin explanation. An executive may need to know whether a competitor is named in the prompts that matter. Each decision needs a different cut of the same raw record.

Begin with questions a buyer could reasonably ask before they know your brand. Sales calls, lost-deal notes, site search and support conversations are better starting material than a list of keywords copied from an SEO tool. Preserve the buyer's constraint. “Best reporting software for agencies that need client workspaces” says more than “reporting software.”

Group prompts by the decision stage they represent, then lock a version of the set. You can add a new prompt when a new market question appears. Do not silently replace an old one, because a trend line made from changing questions is not a trend line about visibility.

• Discovery prompts ask which products or approaches exist.
• Evaluation prompts introduce the condition that changes the choice.
• Comparison prompts name alternatives a buyer is already considering.
• Accuracy prompts test claims about your company, pricing or product limits.

The same prompt can be useful on more than one assistant, but do not assume those assistants are the same surface. ChatGPT search may include citations when it uses the web, and its own help documentation says those citations can be incomplete, outdated or incorrect. That is a reason to save the answer and inspect the source, not to treat a citation as a full retrieval log.

Keep the raw table boring. One row should represent one valid run of one prompt on one surface at one time. A dashboard can summarize later, but the record needs enough detail to survive a question from the person who did not run it.

• Prompt and prompt-set version, including the buyer constraint.
• Assistant, mode, locale and run date, plus any visible model detail.
• Whether your brand was named, recommended, cited or described incorrectly.
• Competitors named and the sources visibly attached to the answer.
• A link to the saved response and an evidence-backed next action.

Keep failed or blocked runs out of the denominator, but show their count. Folding a timeout into “not mentioned” turns an operational failure into a market finding. Likewise, an answer that does not name any product is not evidence that all products lost. It is a third outcome worth reporting.

Mention rate

This is the share of valid runs in which the answer names your company. It is useful for seeing whether you enter the conversation. It does not tell you whether the company was a recommendation, whether the explanation was accurate or whether a buyer could click through.

Recommendation rate

Count a recommendation only when the answer presents your company as a fit for the prompt's stated need. A brand can be named as an example, a warning or an option that does not fit. Counting all three as wins produces an AI visibility metric that flatters the report and misleads the team.

Citation rate

Count this when a source points to a page you own, and record the URL. The page may support a category explanation without the answer recommending your company. The inverse can also happen: a brand is recommended from information that is not visibly cited. Citations and recommendations need separate columns because they suggest different investigations.

Accuracy rate

Where a response makes a checkable claim about your business, score it against a documented fact. Do not call silence accurate. Do not call an attractive answer correct until the relevant owner has verified the claim. The article on wrong business information in ChatGPT explains why that distinction matters.

A mention rate without a competitor record lacks context. If your company appears in four of ten runs, that may be strong if no competitor appears more often. It may be a warning if the same rival appears in all ten with a clearer reason to choose it.

Report the named competitors prompt by prompt. A competitor that dominates broad awareness prompts may be irrelevant to a narrow implementation question. A competitor named only when a buyer asks about a missing capability may be pointing to a product boundary, not a content gap. The report should preserve the distinction before someone creates pages aimed at the wrong rival.

Save the answer text and the visible sources. A label like “competitor cited” is an invitation to infer more than the observation supports. The source may establish a general fact, quote a review or be attached to a nearby sentence with a different claim. Open it before deciding that it caused the recommendation.

Google says its AI features use the same SEO fundamentals as Google Search, not a special AI-only markup or file. That is a useful baseline for a technical check. It does not prove why any particular assistant selected a competitor, nor does it make a published page a promised citation.

Write an action at the confidence level the evidence permits: verify this feature explanation, investigate this comparison question, check crawl access, or contact the publisher of this outdated listing. “Create ten AI-optimized posts” is not an evidence-backed finding. It is a vague project wearing the report's authority.

The summary is not the place to turn uncertainty into certainty. Give the reader the direction of movement, the prompts that drove it, the most material competitor pattern and the next decision. Link each statement back to a saved answer or a page in the detailed report.

For example: “Across the locked evaluation set, Beseen was recommended in six of twenty valid runs, up from three last month. The increase came from agency-reporting prompts. Quarry remained named in every run that required dashboard-layout export. We need to verify whether that reflects a real product limitation before briefing a comparison page.” This is useful because it says what happened, where it happened and what remains unresolved.

Do not claim a percentage-point change caused revenue, a new page or an algorithm change unless the study was designed to isolate that cause. The guide to proving a change moved an AI answer sets a higher bar for that claim.

Use a rhythm that matches the decision and the volatility you can observe. A monthly executive review can make sense for a stable category. A focused check after correcting a false product claim may deserve a nearer follow-up. Re-running a large prompt set every day often produces more noise than learning.

Keep the cadence fixed long enough to compare periods. If the team changes the assistant mode, locale, prompt wording and sample count together, label the period as a method change. A clean break is more honest than a continuous chart whose values are no longer comparable.

Suppose a report says recommendation rate rose from 20% to 30%. The next question is not whether the chart is attractive. Open the runs that changed. Perhaps a new answer began naming the company for a broad question about agency reporting, while all four prompts about client approvals still choose the same competitor. The average improved. The product question that matters did not.

Conversely, a rate can fall because the team added a new family of hard prompts. That may be a useful discovery, not a decline in the old set. Report the locked set and the exploratory set separately. The first answers whether a stable measurement moved. The second helps find the questions the stable measurement was too narrow to catch.

Write a short note beside material changes. Include the changed prompt, how many valid runs produced the pattern, the competitor or source visible in those answers and the reason it matters to a buyer. This makes the chart review faster because nobody has to guess whether a one-point movement came from a high-intent evaluation prompt or a generic question that never influences a deal.

Avoid benchmarks that have no shared prompt set, locale, time window or scoring rule. One vendor's “share of voice” can count brands, citations, answer position or a weighted blend. A comparison of unmatched methods is a comparison of labels, not market position.

Avoid invented precision as well. A decimal score suggests a level of measurement that a small, changing set of generated answers cannot provide. Saying that a company was named in seven of twenty valid runs is more useful than saying its visibility was 35.0, because the reader can see the denominator and ask to inspect the answers.

Do not use the report to turn a competitor's ordinary strength into an accusation. If a competitor is recommended because it supports a capability you do not, say that. The report is there to improve decisions, not to manufacture a content explanation for every loss.

In the first week, collect the current buyer prompts, remove duplicates and label the few that clearly map to active decisions. Run a baseline and save the raw answers. Do not rush into a monthly score before you know whether the prompts describe the business you actually want to win.

In the second week, inspect the losses with the people who own product facts and sales objections. Pick one small action that the evidence supports. That might be clarifying an existing page, checking an incorrect partner listing or researching a comparison page. Record the prediction in plain language: which prompt should this help, and what result would count as progress?

In weeks three and four, finish the work, check technical access where relevant and plan the next comparable sample. The report should not promise that a publication changes every assistant on a deadline. It should leave the team with a baseline, a defensible action and a way to observe what happened next.

The best output is not a larger dashboard. It is a short queue of evidence-backed work: a page that needs a clearer condition, a comparison that needs research, a false fact that needs correction, or a product decision no amount of copy can solve. Give each item an owner and the prompt it is meant to address.

Beseen begins with the prompts, answers, competitors and sources behind an observed gap. Start with a free report when you need that baseline. A report becomes useful when it makes the next piece of work smaller and more defensible, not when it gives everyone another number to admire.