Reading an AI visibility report: share, position, sources, intervals

A category report packs several kinds of number onto one page, and each kind has its own denominator. A count of answers, an average position and a source table are not measured on the same base, so they cannot be compared with each other. This walkthrough takes one public page, camping tents, and reads it line by line.

This page belongs to the October 2026 edition, which is frozen, so it will look the same when you open it: 10 answers, from 5 questions asked once on each of 2 engines. It shows the week of Oct 19, 2026. The 40 US categories updated from November 2 work differently: each rests on 40 answers, and a buying-advice question replaces the alternative-to-a-brand question. The columns and the way to read them stay the same. In this post, "OpenAI" means OpenAI's API with web search (gpt-6-luna) and "Perplexity" means Perplexity's Agent API with the Sonar model (perplexity/sonar).

What is on a category page?

From the top down:

• Header figures. "10 answers sampled" and "11 brands recommended at least once". The first is the denominator for the ranking; the second counts distinct brands.
• The ranking table. One row per recommended brand, with a rank, a count and percentage, an average position, and one column for each engine.
• A badge box. A note that brands recommended in at least 5 of 10 sampled answers can show a badge, with a link to the badge page.
• The sources table. The domains that appeared in each engine's source lists, per engine.
• The five questions. The prompts, each labeled by type, so you can see what was asked.
• A citation line, the page's suggested way to quote its figures with a link back.

Everything below follows that order.

How do you read a count and a percentage?

The page writes each result as a count and a percentage, for example 60% beside "6 of 10 answers". The count is the number of answers, out of all 10 sampled, that presented the brand as a pick. The denominator is every answer for the category: all five questions, both engines. One answer is worth 10 points of share.

The page does not print an interval. The methodology says that a brand's true rate could be anywhere in a wide range and that brands within a few answers of each other cannot be told apart. Here is the range for each count on this page, as the 95% Wilson interval:

Table: the 11 brands recommended at least once on the camping tents page, October 2026 edition, 10 answers, both engines together. Intervals computed from the counts; the page does not show them.

Every interval overlaps every other. The widest gap on the page, 6 of 10 against 1 of 10, still has ranges that meet (31.3% to 83.2% against 1.8% to 40.4%). Overlapping intervals are a cautious test, not proof that two brands are equal, but they do say that a sample of 10 answers cannot separate these counts with confidence. So the useful reading is the shape: eleven brands, four tiers, one brand named in 6 of 10 answers and no other in more than 3. Read it as a short list that one brand appears on more often in this sample, and not as an order of merit.

The badge rule shows the same thing from another side. According to the methodology, a brand earns a badge at 5 of 10 answers, where the lower end of its 95% Wilson interval is at least 20% (13 of 40 once a category has 40 answers). One brand on this page reaches 5 of 10; the other 10 do not.

What does a tied rank tell you?

Rank is by count and nothing else. One brand holds rank 1 at 6 of 10 and one holds rank 2 at 3 of 10. Five brands share rank 3 at 2 of 10, so the next rank is 8, shared by four brands at 1 of 10. A tie is not broken by position or by name; names only set the order in which tied brands are listed. Nine of the 11 brands sit in the two tied groups, so most of the ordering is ties.

How do you check a report's totals?

You can test a page's arithmetic in a minute. Add the counts in the "recommended in" column: 6, 3, five brands at 2 and four at 1 make 23 recommendations. Divide by the 10 answers and you get 2.3 brands per answer. The October edition as a whole averaged 2.9 brands per answer across 1,587 answers, all question types and both engines, so this category is a little below the edition average.

The engine columns add up too: OpenAI's column sums to 12 brand recommendations over its 5 answers, and Perplexity's to 11 over its 5. This works when the page lists every brand. A category page shows at most 15, and the header figure says how many were recommended (11 here, all 11 shown). It is a plausibility check, not proof that the extraction is right. A report whose rows cannot be reconciled with its totals, or that hides the denominator, is harder to trust.

What does average position mean, and what does it rest on?

Average position is the mean place of the brand among the brands an answer recommends, where 1 is first. The page shows it only for brands recommended in 3 or more answers, and prints the number of answers it rests on, for example "1.3 (3 answers)". A brand that an answer only mentions, such as the brand named in an alternative question, takes no place. Position never decides rank.

Here only two brands qualify: the brand at 6 of 10 shows 1.0 over 6 answers, and the brand at 3 of 10 shows 1.3 over 3 answers. The second figure rests on very little. Move one of three answers from first place to fourth and the average shifts by 1.0. Use position only as a description of how those few answers listed the brand, not to order brands or break a tie.

Why are there two engine columns?

Each engine column counts that engine's own answers, 5 each here, so one answer is 20 points. A result of 2 of 5 has a 95% Wilson interval of 11.8% to 76.9%, nearly the whole range. Look for patterns across the columns, such as a brand recommended by both engines, and not for small gaps between them: a difference of one answer is not a difference between assistants.

Engine overlap is, per category, the brands both engines recommended divided by the brands either recommended (a Jaccard measure, not answer-by-answer agreement). With all five questions, OpenAI's 5 answers recommended 10 brands and Perplexity's 5 recommended 7, with 6 of the 11 brands recommended by both: an overlap of 54.5%. Without the alternative question (8 answers), it is 62.5% (5 of 8 brands). The October average over 160 categories, also without the alternative question, was 47.0%, so compare the 62.5%, not the 54.5%. One category is an example, not a finding. The engine overlap post explains the measure.

How do you read the sources table?

The table lists domains that appeared in the source lists of the category's answers, one column per engine, as "x of 5" answers whose list held the domain. The two columns are built differently. For OpenAI, a source list is the pages its answer links. For Perplexity, it is the pages its search retrieved, which the API does not tie to the answer text.

The page states the size difference for this category: OpenAI's lists held 2 domains per answer on average and Perplexity's 16. So a domain at "4 of 5" in Perplexity's column and one at "4 of 5" in OpenAI's column are not the same kind of evidence, and the page orders domains by the average of the two engines' shares, which tends to put Perplexity's domains first.

Table: four of the ten domains listed on the camping tents page, October 2026 edition, counted per engine on 5 answers each.

Two cautions. A domain in a source list shows presence, not influence: nothing here says a page changed which brands were recommended. And a brand's own site, such as bigagnes.com in the table, appears next to retailers and review sites, so look at who is in the list before deciding what kind of gap you have.

What do the five questions change?

The page lists the questions. Four describe a shopper's need: best overall, a budget under $200, a premium option at $500 or more, and a specific need. The fifth, in this edition, asks for alternatives to a named brand. That question names a brand in the prompt, and across the October edition its answers listed 5.0 brands on average against 2.3 for the other four types combined, so it adds more brands, and more counts, to a category page than any one of the other four questions does.

You can see the effect on this category. Without the alternative question there are 8 answers, not 10; 8 brands, not 11; and the most recommended brand is in 4 of 8 answers (50%) against 6 of 10 (60%). Figures that depend on how many brands an answer names, such as brands per category, change with it, so a good report says whether the alternative question is in or out.

What should you check in any AI visibility report?

1. The denominator. How many answers sit behind each percentage, and is the count printed beside it?
2. What was asked. Which question types, written by whom, and does any of them name a brand?
3. Which engines, and how. API or app, with what model and location, and are the engines shown separately?
4. What counts as a hit. A recommendation or any mention? The index counts only recommendations.
5. An interval, or enough information to compute one. Two shares less than a few answers apart are one result.
6. What a source list means. Pages the answer links, or pages the search retrieved? Never add the two.
7. The date and the edition. Which update, and is it like for like with the previous one?

If you want to see how often assistants recommend your own brand, the free snapshot reports it; read it with the checklist above.

What changes when a category has 40 answers?

From the November 2 update, the 40 US categories that are still updated are sampled with 40 answers each. The same 60% share then looks like this:

Table: the same 60% share at both sample sizes, with the 95% Wilson interval. The October column uses counts from the camping tents page; the 40-answer column uses made-up counts of the same size.

The interval narrows from about 52 points wide to about 29, which helps, but brands a few answers apart still overlap. The reading rules above hold at both sizes.

Frequently asked questions

Does a badge mean a brand is the best in its category?

No. A badge records that a brand was recommended in at least 5 of 10 sampled answers in the October edition, or the equivalent count at 40 answers. Rank does not decide it, several brands can hold one, and the methodology says it is not an award or an endorsement.

Can I quote figures from a category page?

Yes, with a link back to the page. The page gives a suggested citation line: the index, the category, the update's date and the page address. Quote the count with its denominator, not only the percentage.