Six Brands, Two Questions: Your Site Gets Cited About Your Price, Not About Your Category

We asked six well known brands two questions each on the same surface on the same day: what does this product cost, and what is the best product in its category. On the price question, all six had their own website cited as a source. On the category question, one did.

That gap is the most useful thing we have measured about which pages are worth writing, and until today we had measured it on one brand.

Disclosure: EchoWi sells AI visibility measurement, so a study concluding that you should measure which questions your site can win is a study with an interest. Here is everything needed to run it against us: the 24 questions are printed in full below, 12 of them in the table and 12 more in the replication, the surface is Google’s AI Overview, the market is the United States in English, each question was asked three times with the cache bypassed, and the dates are 11 and 14 August 2026. Every figure has a row in our measurement register. Anyone can repeat this and publish a different answer.

The short version

1. The price question is one your own site wins. All six brands had a domain of their own cited, and five of the six in every run that answered.
2. The category question is one it does not. One brand of six held a citation on its own domain, and that brand is the exception this article spends a section on rather than the rule it confirms.
3. Being named and being cited are different outcomes. Four brands were named in every run of their category answer while contributing nothing to it. A report that counts those as visibility is counting someone else’s pages.
4. One brand was not even named. Asked for the best email marketing platform, three runs produced an answer that never mentioned the category’s best known name.
5. Where you appear changes as much as whether you appear. The median first mention is at character 1.5 on the price question and character 125 on the category one. On one you are the answer; on the other you are an example inside it.
6. We re-measured our own published claim and it moved. The Shopify pair we have been quoting as three runs of three came back as two of three today. The direction held, the statistic did not, and the section on that is at the bottom.

What was measured

Six brands, each the best known name in its category, and two questions each:

• The price question, which names the brand: “How much does X cost per month?”
• The category question, which does not: “What is the best X?”

Both went to Google’s AI Overview, in the United States in English, three times each with the cache bypassed, on 11 August 2026. For each run we recorded whether the answer named the brand, which domains it was built from, and whether any of those belonged to the brand.

The design is deliberately narrow. One surface, one market, one day, so that the only thing changing between the two halves is what the question is about.

The price question

Six of six had their own domain in the sources. Five of six had it in every run that returned an answer.

Two details are worth pulling out of that table.

Canva answered twice out of three. The third run produced no AI Overview at all, and the denominator here is what answered rather than what was asked for. Reporting Canva as two out of three would turn a surface that declined into a brand that lost, and those are different facts.

The brands often contributed more than one domain. Shopify’s answers cited www.shopify.com, community.shopify.com and help.shopify.com. Squarespace’s cited its main site and its support subdomain. The pages doing the work are pricing pages, help centres and community threads, which is to say the documentation nobody writes for marketing reasons.

The category question

One of six.

The sources these answers were built from are the ones this site keeps finding in every category it measures. “Best ecommerce platform” was built from Forbes, G2, technologyadvice.com and ecomm.design. “Best website builder” from Zapier, Reddit, YouTube, tooltester.com and six more comparison sites. “Best note taking app” from Zapier, Reddit, Quora, YouTube and G2, seven domains and not one of them belonging to a company that sells a note taking app.

Mailchimp is the row to sit with. Three runs, three answers, and the best known name in email marketing appears in none of them. It is the only brand here that failed both outcomes at once: not in the answer, and not in the sources of the answer. Being absent from a list of recommendations is a different problem from being recommended by somebody else’s page, and a dashboard that reports “mentions” collapses the two.

Where you appear, not just whether

The surface returns the position of the first mention, and it separates being the answer from being an example in it.

On the price question the brand is the first thing the answer says, because the answer is about it. On the category question the median brand shows up a hundred and twenty five characters in, and Canva’s first mention in the graphic design answer is at character 411, well past where a reader has decided what the answer said.

This is the part that a mention count cannot express. Two brands both “mentioned in three of three runs” can be the subject of the answer and a footnote to it.

The exception, which is published rather than dropped

Slack holds a citation on slack.com in every run of “what is the best team chat app”. No other brand here manages that, and one clean exception in six is worth more attention than five confirmations.

What we can say about it is small. That answer was built from four domains: Zapier, Reddit, Quora and Slack itself. The website builder answer used eleven, the ecommerce one eleven, the graphic design one eight. A thinner source set with the category leader inside it is what the data shows.

What we are not saying is why, and the temptation to explain it is exactly the one this site has now lost five hypotheses to in four days. “Thin categories let the leader in” is a clean sentence that fits this row and has a sample of one. It would need the same design over categories chosen for how many comparison sites serve them, which is a different study and is not this one.

We re-measured our own claim and it moved

Two articles on this site quote a single measurement as the mechanism behind the rest: asked what the best ecommerce platform is, Shopify holds nothing, and asked what Shopify costs, www.shopify.com is cited in all three runs. That was measured on 10 August 2026 and it was correct.

Today the same pair returns two of three.

The direction is intact and is now much better supported than it was: the category side is still zero, and five other brands behave the way Shopify did. What did not survive is the phrase “in all three runs”, and it did not survive one day.

That is the arithmetic of a three-run statistic rather than a change in the world. With three runs, one miss is thirty three points, so “every run” is the least stable way to describe a citation that is plainly there. Both readings are now in the register with their dates, and the two articles that carried the stronger phrasing have been corrected to say what was measured on each day.

The rule we already had for this is that you re-measure your best number because it is the cheapest one to check and the one that costs most if it falls. This is the first time we have run that against a number of our own that was already published.

Six more brands, and the category half did not replicate

On 14 August 2026 we ran the identical design over six brands this piece had never touched: Calendly, Dropbox, Grammarly, Zoom, HubSpot and Asana, same two questions, three runs each, AI Overview.

The price half replicated exactly. Six of six had their own domain in the sources again, five of six in every run, identical to the first batch.

The category half did not. Three of the five answered category cells cite the brand’s own domain, against one of six here. Pooled across all twelve brands that is 4 of 11 answered cells, so the honest figure is not the near-switch this article originally described. One category question, Zoom’s, returned no AI Overview at all across three runs and is excluded rather than counted as a zero.

It also killed the explanation offered further up. We suggested the single exception sat in an unusually narrow answer, four domains against eleven, and said it needed testing. Calendly is cited on its own domain in an answer drawing on thirteen domains, the widest category cell in either batch, and the two cells that do not cite the vendor include the narrowest. Width is not it, and we are not replacing it with another guess from the same twenty-four cells.

What survives is the asymmetry itself, and it is still large: twelve of twelve on price against 4 of 11 on category. What does not survive is calling it a switch.

What this changes about what you write

Your pricing page, your help centre and your community are the pages that get cited. Every own-domain citation in this study came from that kind of page, on questions that name the brand. They are usually owned by support rather than marketing and are usually the last pages anyone updates.

Your category page probably cannot win the category question. Five of six here did not, and the sources that did win are comparison publishers, forums and video. That is not a reason to write nothing; it is a reason to stop measuring a “best X” prompt as though effort on your own site moves it.

Ask which of your questions has an answer at all before either. 2 of the 24 measurements here returned no AI Overview on one of their runs and 1 returned none at all, and in categories we have measured elsewhere the surface does not draw at all. Optimising for a box that is not being rendered is the most expensive mistake available.

And separate the two outcomes in whatever you report. Named and cited came apart in four of the six category measurements. If a tool gives you one number, ask which of the two it is counting.

What this does not show

One surface. Google’s AI Overview. ChatGPT, Gemini and AI Mode were not part of this design, and we have published elsewhere that source lists do not transfer between surfaces.

One market and one language. United States, English. Our own cross-market work says the ordering of categories travels and the specific domains do not, so the six sources named above should be expected to be different names in Spain or Germany.

One day. The Shopify row in this study is itself the argument for that caveat.

Six brands. Six is enough to retire a claim built on one and not enough to put a rate on it. What is published here is a count, not a percentage, for that reason.

Nothing here is causal. These are observations of what a surface returned. No intervention was made on any of these sites and none of these brands is a customer of ours.

Common Questions About Being Cited by AI

Why would an AI cite a comparison site rather than the vendor it is describing?

Because the question asked for a comparison and a vendor’s own pages do not compare. A pricing page describes one product, which is exactly what the price question needs and exactly what the category question does not. In this study the same six sites that were cited about their own price contributed nothing to their own category’s answer on the same day.

Is being named in an answer worth anything if my site is not cited?

It is worth something and it is worth less, and the difference is where it comes from. A mention with no citation of yours rests entirely on what third parties published about you, so it moves when they update their comparison and you find out afterwards. A citation is the half you can act on directly. Any report that gives you a single visibility number is adding those two together.

Should I stop trying to rank for “best X in my category”?

Not stop, but stop expecting your own domain to be the source. Five of the six brands here are the best known name in their category and none of their sites was cited on it. The work that moves a category answer is getting into the pages that are cited, which is a placement problem rather than a publishing one, and it is worth knowing that before a quarter of content is committed to it.

How many runs do I need before a citation figure means anything?

More than one, and enough that a single miss is not a third of your statistic. This study uses three per question and one of its rows still moved from three of three to two of three in a day. Answers vary between identical requests, so a single check cannot tell a stable citation from a coincidence, and a three-run “every run” is the most fragile way to phrase one that is really there.

Does this apply outside software?

Unmeasured here, and worth checking rather than assuming. The six categories are all software or design tools, chosen because the brands are unambiguous category leaders. We have measured in local services that the surface often draws no answer at all, and in education that the stable sources are accommodation and guidance platforms rather than institutions, so the shape of who holds the sources changes a great deal by sector.

What is the cheapest useful thing to do with this?

Ask your own two questions and compare them. Run the price question that names you and the category question that does not, three times each with the country and language fixed, and list the domains each answer was built from. If your site appears on the first and not the second, you have the same result as six well known brands, and you know which page to improve and which to stop counting.