Being read and being chosen are two measurements, and no single reading reports both
A citation count tells you which pages an answer was assembled from. A blind naming scan tells you which business the answer put forward. They have different denominators and must never be added.
Published by AI Knows Us · Updated 19 October 2026
Two questions get asked as though they were one. Which pages did the answer come from, and which company did the answer recommend. Our own records answer them separately, with separate instruments, separate denominators and separate dates, and they cannot be combined into a single figure without producing a number that describes neither.
The citation side of our records is a run of 1,260 readings taken from ChatGPT and Gemini on 16 October 2026, holding 20,143 citations across 4,947 distinct domains, counted as one domain in one reading and deduped. That instrument answers the first question and is silent on the second. The naming side is a separate scan on different apps, described below in its own section, and it answers the second question only.
What a citation count can and cannot settle
| Question | A citation count |
|---|---|
| Which pages was this answer assembled from | Answers it directly, domain by domain |
| How crowded is the field for this question | Answers it: 4,947 distinct domains behind 20,143 citations |
| Which kinds of page get used | Answers it: 82.0% of those citations landed on a vendor's or an agency's own site |
| Which company did the answer recommend | Does not answer it at all |
| Was a business named in the text | Does not answer it at all |
The last two rows are not a weakness in the instrument. They are what the instrument is not pointed at. A domain can be in the source list of an answer that never names its owner, and a business can be named in an answer whose source list does not contain its site.
The naming side, kept in its own paragraph on purpose
Our separate blind scan was run on 2026-10-16 on ChatGPT and Perplexity, against clawlaw.in, a legal search company built by our founder. The questions never contain the company name. 6 of the 10 answers named it. Two of those readings are published in full, and they have to be read as a pair.
- On a pricing question for a solo practitioner advocate in India, it was named first on ChatGPT alongside four other businesses, with 5 sites read, and named first on Perplexity alongside three others, with 5 sites read.
- On a question about the most comprehensive court record search across Indian courts and tribunals, it was not named on ChatGPT, which named no business at all and read 6 sites, and not named on Perplexity, which named two publishers and read 8 sites.
One question won and one question lost, on the same domain, on the same day, on both apps. Quoting the first without the second would turn a measurement into an advertisement.
The rule, stated plainly
Those are four datasets in this programme and these two sections quote two of them. They sit on different apps, on different dates, with different denominators, and a sentence that slides from one to the other is the most expensive error available in this field. So:
- Never add them. A citation count and a naming count do not sum to anything.
- Never average them. There is no shared unit to average across.
- Never present one as evidence for the other. They fail for different reasons: a page can be unreadable, or readable and unconvincing.
- Name the dataset in every sentence. Which apps, which date, which denominator. Without that label, a figure cannot be checked and should not be quoted.
Why you need both columns anyway
Because the two halves of the thing a business actually wants are being read and being chosen, and they break independently. A business absent from every source list has a reach and evidence problem, which is answered by publishing pages that can be fetched and quoted. A business present in source lists and still not put forward has a different problem, which is that the pages being read do not contain the thing that would settle the choice.
The diagnosis is only possible with both columns side by side and labelled. One column alone tells you that something is wrong without telling you which half it is in, and the two fixes cost different amounts and take different lengths of time.
What to put in your own report
- One column for sourced readings, with its own denominator, per app.
- One column for readings that named your business in the text, with its own denominator, per app.
- A third column recording readings that returned no sources at all, because those cannot be won with a page and belong outside both rates.
- The date on every column, and the date of the next reading set before the result is seen.
- No combined score anywhere on the page.
Running that pair of readings on a schedule, and writing the pages that move the first column, is what we sell. The Google Search Console export behind our own separate search side record, with its window and export date printed, is on our proof page.
Questions about citations against recommendations
How big is the citation side of your data?
1,260 readings on 16 October 2026 across ChatGPT and Gemini, holding 20,143 citations across 4,947 distinct domains, counted one domain per reading and deduped.
How big is the naming side?
A separate blind scan of 10 answers run on 2026-10-16 on ChatGPT and Perplexity, in which 6 of the 10 named the subject. Different apps, a different question set and a much smaller size, which is why it is never merged with the citation run.
Can I convert one into the other?
No. They count different events with different denominators, and a ratio between them would be arithmetic performed across two unrelated instruments.
Which one should I measure first?
Measure both from the start, because the first reading of each is what every later reading is compared against. If only one is possible, the naming column is closer to the commercial outcome and the citation column is closer to the cause.
Where is the method for the citation side?
The earlier run in the same programme, with its question design, controls and complete source mix, is published as 597 readings across ChatGPT and Gemini in three markets.
Where the confusion comes from
It is not carelessness. The two events look identical from the buyer's chair, because a buyer reading an answer sees one block of text with some links under it and has no reason to separate the two. The separation only becomes visible when you try to fix a result, and at that point it decides everything about what gets worked on.
Vendors also have an incentive not to separate them, because the larger of the two numbers makes a better slide. A citation count is usually the larger one, and it is also the one further from the sale, so a report leading with it and implying the other is making the most flattering available substitution. Asking which of the two a figure counted, every time, costs one sentence and settles it.
Two columns, two denominators, one date each
Every misleading number in this category comes from the same move: two measurements with nothing in common, printed as one score. Keep them apart in the spreadsheet, keep them apart in the report, and keep them apart in the sentence. A report with two labelled columns is harder to write and impossible to misread, which is the trade worth making.