We Asked AI 40 SEO Tool Questions. Two Brands Answered 90%.

Across 40 buying questions put to a search-grounded AI assistant, Semrush appeared in 92% of answers and Ahrefs in 82%. Thirty-nine of the forty answers named one or both. If you sell an SEO tool and you are not one of those two, AI assistants are largely not recommending you.

We ran this because we sell AI-visibility tracking, and it seemed only fair to point the method at our own category first. UtilitySEO appeared in zero of the forty answers. That result is published here unchanged.

What we did

Forty buyer-intent prompts, written in advance and never revised, covering the questions people actually ask when choosing a tool: best audit tool, cheapest alternative to Semrush, best for a JavaScript site, what detects cannibalisation, what tracks AI citations, and so on. Each was sent once to Gemini 2.5 Flash with Google Search grounding enabled, in September 2026.

Every answer was scanned for 38 known SEO brands using case-insensitive word-boundary matching, so "Moz" does not match "Mozilla". No model was asked to grade another model. Every raw answer was retained so the counts can be audited.

The results

Share of voice is the percentage of the 40 answers in which a brand was named at least once.

Thirty-four distinct brands were named at least once, and the median answer listed eight tools.

Five things the data shows

1. Recommendation is far more concentrated than search. The top five brands took 48% of all 306 brand mentions. In organic search results for the same queries you would typically see a much longer tail of sites competing. AI answers compress the field.

2. Being the default beats being the best fit. Semrush and Ahrefs appear in answers to questions they are a poor fit for, including "recommend an SEO tool for a solo founder on a budget". The models reach for the most-discussed brand rather than the most appropriate one.

3. Free and first-party tools rank remarkably well. Google Search Console appeared in 72% of answers and Google Analytics in 38%, without either being a commercial SEO product. If you sell a tool, your real competitor in an AI answer is often something free.

4. The long tail is very thin. Fifteen of the 34 brands named appeared in under 10% of answers, and three appeared exactly once. There is a steep cliff between the recognised names and everyone else, with little middle ground.

5. One question was open. Of the forty prompts, exactly one produced an answer naming none of Semrush, Ahrefs or Moz: "Which SEO tool is easiest for beginners?" That is the only prompt in the set where the incumbents did not appear at all.

What this means if you sell software

Being absent from AI answers is not a content problem you can write your way out of. These models are drawing on how often a brand is discussed across the web, in reviews, comparisons, forums and news, rather than on how well any single site is optimised. Publishing more pages about yourself does not move it.

What plausibly does move it is the same thing that has always built authority: being mentioned by other people, in places that get read. Third-party reviews, comparison round-ups written by someone else, and original data that other sites cite. Note that a brand does not need to be linked to in order to be named in an AI answer, which makes unlinked mentions more valuable than they used to be.

What this means if you are buying a tool

Treat an AI recommendation as a starting shortlist, not a verdict. The data suggests the answer you get is weighted heavily toward the largest brands regardless of your stated constraints, and that smaller tools which might fit your budget or use case better are frequently not surfaced at all. Ask follow-up questions that force specificity, and check whether the tool named actually does the thing you asked about.

Limitations, stated plainly

This is one model at one point in time. Gemini 2.5 Flash with Google Search grounding was the only assistant sampled, so these numbers describe that system and not ChatGPT, Claude or Perplexity, which draw on different sources. Answers are non-deterministic and would vary somewhat on a re-run. All prompts were in English with no location set. Each prompt was asked once rather than repeated for an average.

Counting a brand mention is not the same as counting a recommendation: a tool named in passing as a comparison point counts the same as one recommended enthusiastically. A larger study would separate the two and sample every major assistant repeatedly over time.

The data

The full prompt list, every raw answer and the counting code are available on request. We would rather people check the numbers than take them on trust, and if you re-run it and get materially different results we would like to know.