Semrush found comparison queries name a brand 2.4x as often: the page to publish next

On 9 June 2026 Semrush published a study of 3,981 domain appearances across 115 prompts and four AI engines, and the line I keep coming back to is this: comparative queries produced a 43.3 percent brand mention rate against 18 percent for informational ones, which the study rounds to 2.4x.

I read it the week it came out and drafted a sample comparison page for a dental clinic query the same afternoon.

The first draft was an ad.

It compared the clinic with two names ChatGPT already gave for the same query, and every row leaned our way, and I knew a reader would smell it in ten seconds.

So I threw it out. I did not know if the honest version would work.

This post is what I would build now, in the order I would build it, with every number attached to the page it came from.

Play 1: What the data says, and what it measured

The Semrush ghost citations study, fetched 25 September 2026, found that 61.7 percent of AI citations never named the brand in the answer. Cited, linked, and invisible.

But the split by query type is the useful part. Informational prompts such as "what is" or "how does" got an 89.3 percent citation rate and an 18 percent mention rate in the Semrush data, while comparative prompts such as "best", "vs" and "recommend" got a 43.3 percent mention rate.

So the 2.4x in the Semrush write-up is a ratio of mention rates between query types. It is not a measurement of pages. Nobody built two pages and counted.

And that matters for how far you push it. The honest reading is that assistants name brands when the question asks them to choose, and a page built for a choosing question is the page most likely to be in the room when that happens. That is a strong prior, not a controlled result, and I would rather say so here than have you find out from a competitor.

The same week I checked the other correlation everyone quotes.

The Ahrefs brand-mention study put the correlation between brand mentions and AI citation at r=0.66, against 0.22 for backlinks. Mentions on other pages, not links to yours.

A comparison page is the one page format where those two things overlap on purpose.

Play 2: What a comparison page is, and is not

Ask an assistant the question your customers ask, and write down the two names it gives.

Then build one page that compares you with exactly those two, on the criteria a customer would use, with the rows filled in for all three.

That is the whole format. Three columns, one of which is you.

It is not a "best clinics" listicle with you at number one. And it is not a page that names competitors only to knock them. Google's page on creating helpful content asks whether a page offers "original information, reporting, research, or analysis" and whether it exists for people or for search engines, and a page written to flatter its owner fails both questions on sight.

Praise your competitor first.

I mean that literally.

Open the page by stating what each of the other two does well, in one plain sentence each, before you say a word about yourself.

It reads as fair, which is the only reason anyone finishes the page, and it gives the assistant three named entities with attributes instead of one brand and two strawmen.

The rewritten clinic page leads with the competitor that has the longer opening hours. The clinic does not, and saying so costs nothing.

Play 3: The table format that gets lifted

Assistants lift tables because tables are pre-structured answers. But the structure has to be honest or it is worthless.

Here is the layout I use at NovaTechRay for every local comparison page, criteria down the side, names across the top:

Three rules keep it clean.

Every cell for a competitor comes from their own public page, and you link that page. Where they publish nothing, the cell says "not published", which is both true and quietly damaging.

And you include at least one row where a competitor wins.

Then add a dated line under the table saying when you last checked the other two columns. A stale comparison is worse than none, and the date tells both reader and machine that someone is maintaining it.

I did not know if the "not published" cells would read as petty. They are the cells a customer most wants filled before they call, so I stopped worrying.

Play 4: Where the price line goes on it

Put your price in the first row of the table and again in the first 200 characters under the H1, which is where Ahrefs says the answer should sit.

The reason is a SIGIR 2026 paper by Vishwakarma, Kumar and Jamidar, arXiv 2605.25517, fetched 25 September 2026. Across 252,000 trials on six models, the SIGIR paper found a page carrying an explicit price beat the identical page without one at an odds ratio of at least 6.26, and on two models the effect went past the measurement ceiling.

The SIGIR paper puts formatting-only changes between 0.78 and 1.68. So the table helps because of what is in it, not because it is a table.

I wrote up the full paper, with its limits, in the research says hiding your prices keeps you out of AI answers. The short form: the testbed injects two documents and asks which gets cited first, so it measures what wins a comparison you have already entered, not how you enter one.

But that is exactly the situation a comparison page creates. You have put yourself in a three-way slate on purpose. The price row is the cheapest thing on the page and the one with a controlled study behind it.

If you cannot publish a single figure, publish a from-price with what moves it. The paper's factor is whether price information is present, not whether it is precise.

Play 5: Two mistakes that turn it into an ad

The first mistake is winning every row.

Nobody wins every row.

A reader who sees a clean sweep assumes the criteria were chosen backwards from the result, and they are usually right. Pick the criteria a customer would use, including the ones you lose, and let the table stand.

The second mistake is comparing yourself with names the assistant never gives.

It is tempting to build a page against the two competitors you resent most, even when neither one appears in a single answer. That page compares three options the assistant is not choosing between. Ask the question first, in the assistant your customers use, and compare against whatever it says, even if the answer annoys you.

There is a third one I nearly made.

My first plan for the clinic page was to ship it as a landing page with no entry in the site navigation. A crawler would have had no path to it.

Every rule I have written about being cited but never named assumes the page is fetchable in the first place, and I almost skipped my own step.

So does the honest version work? I am not going to put a number on a single page here. The Semrush split says the format points the right way, the SIGIR paper says the price row pulls its weight, and that is the honest limit of what I can claim.

The short version

Ask the assistant the customer's question and write down the two names it gives.

Build one page that compares you with exactly those two, on criteria a customer would use.

Praise each competitor in one honest sentence before you mention yourself.

Put your price in the first row and in the first 200 characters under the H1.

Include at least one row you lose and mark unknowns as "not published".

Date the table and link it from the navigation.

This is what we do at novatechray.com.