0.8% then 0.7%: the review directory finding has now been measured twice
Two question banks designed eleven days apart put review directories at 0.8% of 10,652 citations and 0.7% of 20,143, which is as close to settled as this field gets.
Published by AI Knows Us · Updated 19 October 2026
Review directories carried 0.8% of the 10,652 citations recorded across 597 usable readings on 15 October 2026, and 0.7% of the 20,143 citations recorded across 1,260 readings on 16 October 2026. Both runs read ChatGPT and Gemini. The second question bank was written deliberately in eight directions the first one never went, and it came back with the same answer to a tenth of a percentage point.
One measurement of a share like that is a number. Two measurements from independently designed banks is a replication, and a replication is the part worth acting on, because it is the thing that will still be true when the wording of your own questions differs from ours.
What makes this a replication rather than a repeat
A repeat asks the same questions again and tests whether the engines answer consistently. A replication changes the instrument and tests whether the finding survives. The second bank changed almost everything about the instrument.
| Field | First run | Second run |
|---|---|---|
| Date | 15 October 2026 | 16 October 2026 |
| Questions per market | 102, being 72 identical across markets plus 30 market specific | 210, in 21 batches |
| Readings | 597 usable of 612 attempted | 1,260, all usable |
| Citations recorded | 10,652 across 2,708 domains | 20,143 across 4,947 domains |
| Review directory share | 0.8% | 0.7% |
| Design intent | Breadth across three markets | Eight directions the first run missed, including procurement depth and competitor named questions |
The second bank reached into price and procurement depth, questions that name a competitor, role and seniority, the problem stage before the category is known, comparison with adjacent search suites, surface depth, vertical depth and time. Several of those are exactly the shapes of question a directory is supposed to win. The share did not move.
What a directory listing is still good for
Three things, none of which is a citation, and all three of which are reasons to keep a profile current.
- People read them. Human beings comparing suppliers open these pages on purpose, and nothing an engine does to its source list changes that visit.
- They settle facts. An accurate listing with your current coverage, your current price band and your current contact details is one fewer place for a stale fact about you to live.
- They are evidence for a human. A buyer who wants a second opinion about a supplier can get one there, which is a job your own site cannot do for itself.
What changes with these two numbers is only the line of the budget the profile sits on. It belongs with the things you buy to reach buyers, and not with the things you buy to be cited.
How to check this on your own category rather than borrowing ours
Our bank is about our own category and yours will differ. The test is cheap and the result is yours.
- Take thirty to a hundred questions in your buyers' own words, with no company name in any of them, including yours.
- Run them on two named apps, record every cited domain, and keep one column per app.
- Classify each cited domain once: own site, official documentation, news, community, directory.
- Print the share for each class with the denominator beside it, and the date.
- Repeat it with a differently written bank before you believe any share you get.
That last step is the one almost nobody takes and it is the one that turns a figure into a finding. It is also what makes a published rule about spending defensible later, in the same way a written policy outlasts a verbal one, which is the argument behind a marketing policy template you can copy section by section.
Where both readings stop
Two apps, ChatGPT and Gemini, on two single dates. Three further surfaces are part of our product and part of neither run. Neither bank holds a per industry rate, because the vertical questions named no industry. City level figures come from a separate scan of ten named places. And both shares count a domain in a source list, which is not the same event as a business being named in the text of an answer, a distinction worth holding onto whenever a directory tells you about its reach.
We sell this measurement and the publishing work that follows, which is why the uncomfortable row gets printed alongside: our own domain was in the source list of 3 of 597 readings and then 6 of 1,260. The search console side of our own record, both windows with the export date, is on our proof page.
What a 0.7% share is not evidence of
Three claims could be built on these numbers and none of them is supported, so they are worth naming before somebody else makes them.
- It does not show that collecting reviews changes an answer. Both runs are source counts rather than experiments. Testing cause needs the same questions re-asked over time while reviews accumulate, which is a different design and a longer one.
- It does not say directories are absent from what a model already holds. A share measured on source lists describes the fetched part of an answer and nothing about the rest.
- It does not transfer to another category. Our bank asks about our own market. A category where the recognised directory is a sector body rather than a global platform could sort very differently, and the only way to know is to count your own.
Keeping those three separate from the finding is the difference between a number a reader can use and a number that gets quoted in an argument it cannot settle.
Questions about review directories and AI citations
What share of citations went to review directories?
0.7% of 20,143 citations on 16 October 2026, and 0.8% of 10,652 on 15 October 2026. Two differently designed banks, a tenth of a point apart.
Should I cancel my directory subscriptions?
Judge each one on whether buyers in your category actually arrive through it, which is a question your own analytics answers. What these two numbers settle is that it is not a citation strategy.
Did any directory reach your top ten?
In the 15 October reading, one directory domain was cited in 49 of 597 readings, which is real presence in a small category. In the 16 October top ten there is no directory at all.
Why does a replication matter more than a bigger sample?
A bigger sample makes one instrument steadier. A second instrument tests whether the finding belonged to the first one. The second bank here was built to go where the first did not, so agreement across them is evidence about the channel rather than about the wording.
Which engines were read on both dates?
ChatGPT and Gemini on both. The first run's full method, question design and source mix are published as 597 readings across ChatGPT and Gemini in three markets.
The number that survived a different instrument
Most figures in this field come from one run, by one party, on one bank of questions, and they move the moment anybody asks differently. This one did not. Keep the profiles for the buyers who read them, take them off the citation scoreboard, and spend the argument you were going to have about directories on choosing which of your own questions to answer first.