AI Visibility Competitor Benchmark: How to Compare Against Like-Sized Brands

AI engines cite big, well-known sources far more than everyone else. A study of AI Search Arena data (arXiv:2507.05301, March to May 2025) found the top 20 news sources accounted for 67.3% of all citations for OpenAI models. Its concentration score (Gini) was 0.83 for OpenAI, 0.77 for Perplexity and 0.69 for Google.

The same pattern shows up in brand data. Semrush's index of 126 million prompts found 36 brands visible across all four surfaces it tracked every month, names like YouTube, Google, Reddit, Amazon and Walmart. That finding is mostly a measure of brand size.

The controlled research removes size on purpose. The 252,000-trial factorial anonymised brands by design and warned that real systems may favour established domains. You can read more in our summary of the 252,000-trial GEO factorial.

What this means for you: if you benchmark a 30-person company against the category leader, the gap mostly reflects fame. You can't fix fame this quarter. A like-sized comparison set strips out much of that effect, so whatever gap remains is more likely to point at something you control.

You can't control for brand prominence perfectly. Nobody can, because the common "controls" (Ahrefs DR, Moz DA) are vendor regressions built from backlinks and traffic, which are proxies for prominence itself. So pick peers on plain business traits instead.

1. List every brand a buyer would shortlist alongside you. Use sales call notes, lost-deal reasons and review site category pages.
2. Remove the category giants. If a brand is a household name or many times your size, take it out of the core set. Keep one as a reference line if you like, labelled as such.
3. Match on observable traits. Same product category, same customer type (SMB, mid-market, enterprise), same main geography, similar stage (bootstrapped, seed, growth) and a similar price band.
4. Aim for four to six peers. Fewer and one outlier dominates. More and the set drifts back toward bigger brands.
5. Write down why each peer is in the set. When someone asks why a rival was excluded, you have the reasoning on file.
6. Freeze the set for at least a quarter. Swapping peers every month makes trends meaningless.

Say you run a hypothetical 40-person payroll tool for restaurants. Your set is four other restaurant-focused payroll vendors at a similar stage, not the national payroll providers every accountant already knows. For a wider view of the method, see how to compare your AI visibility against competitors.

Consistency matters more than volume. The same prompts, run the same way, beat a large prompt set run carelessly.

1. Write 15 to 30 buyer questions. Use the questions a buyer asks before they know your name: "best payroll software for restaurants with tip pooling". Keep a few that name a competitor, and track those separately. Our page on tracking prompts that name competitors covers when that helps.
2. Run every prompt on every engine you care about. Score ChatGPT, Gemini, Google AI Overviews, Perplexity and Claude as separate columns. Google says AI Overviews and AI Mode "may use different models and techniques", so treat them as distinct surfaces.
3. Run logged out, with memory off. Saved memories and custom instructions change answers, so logged-in runs aren't reproducible.
4. Record whether the answer cited sources. Retrieval is discretionary on every platform. An answer with no citations may come from the model's frozen training data, which nothing you publish this quarter can change.
5. Log mentions and citations separately. Being named in the text and being linked as a source are different results. See citations vs mentions.
6. Note what each answer says about each brand. Wrong prices, old product names and confused competitors are findings in their own right.

The free checker on this site gives a quick snapshot for up to three buyer questions if you want a starting point before building a full prompt set.

A gap tells you where to look. It doesn't tell you why. A 2026 critical survey of 45 GEO studies (arXiv:2607.14035) found no technique with a stable, longitudinal, cross-platform causal effect on discoverability. So if a rival appears more and you notice they use some tactic, that's a coincidence until shown otherwise.

Schema is the classic trap. Ahrefs' matched-control test of 1,885 pages found AI Overviews citations fell 4.6% after adding schema. A competitor's markup isn't the reason they appear.

Work through gaps in this order, from things you control to things you don't:

The Otterly 2,000-blog experiment shows why the per-engine view matters: Microsoft Copilot produced 69% of one site's tracked citations after Google had deindexed it. A blended score would have hidden that. To check the access row quickly, 99Visibility's free AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.

For the third-party row, our comparison of your own site vs third-party mentions explains where each route leads. For prompt-level detail, use the competitor prompt gap analysis.

A good benchmark

• Peers are close enough in size that no single brand wins every prompt on every engine.
• Each engine has its own row, and the differences between engines are visible.
• Gaps come with a named cause to check, such as "Perplexity: PerplexityBot blocked at CDN".
• Wrong facts about your brand are listed, with the correct fact and the page that should state it.
• Month-to-month changes are small and explainable.

A bad benchmark

• The comparison set includes a category giant that wins everything, so every gap reads as "be more famous".
• One blended score across all engines, with no retrieval state recorded.
• A gap is explained by a tactic the competitor uses ("they have an llms.txt") with no check of access or indexing.
• The prompt set or peer list changed since last month, so the trend compares different things.
• The report claims an AI "rank" or "position". Google's AI reporting exposes impressions only, so there is nothing to validate a position against.

If you want a sense of what a number means in isolation, read what a good AI visibility score is. Inside a like-sized benchmark, the relative gap to peers matters more than the absolute score.

• Comparison set of four to six like-sized peers, with written reasons for each
• Category giants excluded from the core set (or kept as one labelled reference line)
• 15 to 30 buyer questions, frozen, with competitor-name prompts tracked separately
• Every engine scored as its own column, AI Overviews and AI Mode kept apart
• Runs done logged out with memory off
• Retrieval state recorded (cited sources or no cited sources)
• Mentions and citations logged separately
• Wrong facts about each brand noted, with the correct value
• Each gap paired with a check: robots.txt per bot, CDN or firewall response, Google and Bing index status, JavaScript rendering, third-party mentions
• No gap attributed to schema, llms.txt, content style or authority scores
• Date of the run and any known engine changes noted

Run the full benchmark once a month. AI answers vary from run to run, so a single week's result is noise. Monthly runs with a frozen prompt set and frozen peers give you a trend you can trust.

Re-run outside the schedule when something changes:

• You changed your site. A redesign, a move to client-side rendering, or a new robots.txt can remove you from an engine overnight.
• Your CDN changed its defaults. From 15 September 2026, Cloudflare blocks Training and Agent crawler categories by default on ad-displaying pages for new domains and existing free-tier customers. That can change access without anyone editing a file.
• An engine changed models or defaults. As of October 2026, consumer apps switch default models often, and the API model you test may not match what the app serves.

Refresh the peer set once a quarter. Keep the old results, but mark the break so you don't compare one set against another.