Who should you benchmark against in AI answers? Picking a rival set that means something

Why your SEO competitors are the wrong benchmark

Your AI visibility benchmark should use the brands an engine names when your buyers ask, and that list rarely matches your search rivals. In search, a competitor is any site that ranks for your keywords: a publisher, a marketplace, a glossary page. In an AI answer, the competitor is a brand the engine recommends by name in the same answer as you, or instead of you.

The difference shows up as soon as you run real prompts. A payroll software firm may compete in search with HR blogs and government guidance pages. In an AI answer to "best payroll software for a 50-person company", the engine names four or five vendors, and some of them may never rank on the firm's keywords.

Our view: a share-of-answer number is only as meaningful as the set it is divided by. Pick rivals the engines do not name, and your share looks healthy while the brands that win the answers go unmeasured. Pick a set that changes every month, and the trend line measures your spreadsheet rather than the market.

How concentrated is your category?

It depends heavily on the sector, and that changes how many rivals you need to track. Semrush, which sells search and AI visibility software, analysed 126 million US prompts across ChatGPT, Gemini, AI Mode and AI Overviews from January to April 2026. The top three brands held 82.9% of AI visibility in news and media and 76.9% in consumer electronics.¹ In finance the top three held 41.4%, and in industrial 42.2%.¹

In a concentrated category, three or four rivals capture most of the answers, and a small set tells the story. In a fragmented one such as finance, the remaining share is spread across many brands, so you need more names to see who is gaining.

Engines also differ in how many sources they draw on. In the same Semrush data, ChatGPT cited about 15 sources per response against about 3 for Gemini.¹ An engine that cites more sources has room to name more brands, so a rival who is invisible in Gemini may appear in most ChatGPT answers.

Search competitors vs answer competitors

Search competitors

• Any site ranking for your keywords
• Includes publishers, glossaries, marketplaces
• Found with a rank tracker
• One list for every engine

Answer competitors

• Brands named in the same answer as you, or instead of you
• Vendors and providers only
• Found by running your buyer prompts
• Checked engine by engine and country by country

Rivals and sources are two different lists

Split what you find into rivals, which you compete against, and sources, which you compete to be cited by. A review platform, a trade publication or a marketplace often appears in answers in your category. It is not a rival. It is a page that decides whether you get named.

The split matters because sources vary by engine and by sector. Profound, which sells AI visibility software, counted 11.84 billion citations across eight engines from 16 April to 16 July 2026. About 57% went to company-owned sites, with 69% for Gemini and 47% for ChatGPT.² By sector, the median brand-citation share ran from 74% in cybersecurity down to 15.9% in government and nonprofit.²

"Company-owned" means any company's site, so in cybersecurity most citations go to vendors, which makes rivals' own pages the thing to beat. In a category where most citations go to third parties, the source list does more work than the rival list. We covered the engine split in our piece on brand sites' share of AI citations by engine.

How often an engine cites at all varies too. Similarweb, which sells web analytics, found that the share of US ChatGPT prompts whose answer carried a citation rose from 1.6% in June 2025 to 6.8% in May 2026. Travel ran at about 23% and professional services under 4%.³ Our note on citation rates by sector covers what that means for each industry. In a low-citation sector, the brand names in the answer text matter more than the links under it.

How to pick five to eight rivals per country

Choose from runs, not from memory. The method below takes a prompt set you already have and turns the names in the answers into a fixed list.

1. Run the prompt set first. Use 30 to 60 prompts written from real buyer questions, in each country you sell in, across the engines your buyers use. Run each prompt on more than one day, because answers vary between runs.
2. Extract every brand named in the answers, and count in how many prompts each one appears.
3. Remove sources: marketplaces, publishers, review sites, government pages. Move them to the source list.
4. Rank the remaining brands by how often they appear. Take every brand above a threshold you set in advance, such as one prompt in ten.
5. Add the category leader if it is not already there, and any direct rival your sales team names that the engines miss. Mark that rival as "not yet named", because its absence is itself a finding.
6. Cap the list at eight. Beyond that, the extra names rarely change the decisions you take.

Do this per country. A rival that dominates answers in the United States may be absent in Germany or India, where local firms take its place. One global set hides exactly the differences a country team needs to see.

For reference, our own audits run each prompt on three separate days across six engines (ChatGPT, Gemini, Perplexity, Claude, Copilot and Google AI), using official APIs where they are offered and a licensed AI-answer data provider otherwise. The audit page describes how the rival set is agreed before the first report.

A worked example

Here is the method applied to an invented firm. Northwind Digital sells project-management software in the United States and Germany. Its marketing team listed six rivals from memory. The runs told a different story.

Illustrative. Northwind Digital is fictional and the counts are invented round numbers.

Two things changed. Rival C, a German vendor nobody on the US team had listed, turned out to be the main threat in Germany. And the review marketplace, which the team had treated as a competitor, became a source to work on: getting listed and reviewed there is how Northwind gets named.

Keep the set stable, and change it by rule

Freeze the set for at least a quarter. Share of answer is your mentions divided by all mentions in the set, so adding a strong rival mid-quarter lowers your share even if nothing changed in the answers.

Write the change rules down before the first report. A brand that crosses the threshold in two consecutive monthly runs joins at the next quarter. A brand that falls below it for a full quarter leaves. When the set changes, restate the previous quarter on the new set so the trend stays comparable.

Our view: report the rival set on the first page of every visibility report, with the date it was fixed. A buyer reading the number should know what it is divided by. Our buyer's spec for an AI visibility score lists the other things that report should state.