How to benchmark your brand's AI search visibility

AI search visibility benchmarks are reference values for how often engines name your brand and link your domain, across ChatGPT, Perplexity, Gemini, and Google AI Overviews. Share of voice is those mentions against tracked competitors. Good means the top quartile of your vertical on one consistent denominator, across dozens of prompts and repeated runs.

What are AI search visibility benchmarks?

AI search visibility benchmarks are reference values for comparing your mention rate and citation rate with peers. Compare share of voice as well to judge whether your brand’s presence in AI answers matches peers or falls above or below them. They come from vendor and agency panels that run thousands to millions of prompts through the engines and record which brands get named and which domains get linked. The engines in scope for most B2B programs are ChatGPT, Perplexity, Gemini, and Google AI Overviews, with Google AI Mode added by teams whose buyers use it. Teams also add Claude and Copilot when their buyers use them.

A benchmark is only useful if its denominator matches yours. A citation median computed across six engines describes a different event from a ChatGPT-only citation rate. An appearance rate across eight platforms differs from both, so keep their denominators separate when reading the B2B SaaS figures later in this article. Treat every benchmark as a definition first and a number second.

What metrics sit behind every AI visibility benchmark?

Use four metrics on one response set: mention rate, citation rate, share of voice, and a composite visibility score that accounts for prominence. Fix that response set, your prompt library run on each engine a fixed number of times inside a fixed window, to compare the metrics week to week and brand to brand.

What is AI share of voice?

AI share of voice is your brand’s mentions divided by the total mentions of all tracked brands in the same response set, expressed as a percentage. If your category prompt set produces 400 brand mentions across four engines in a week and 60 of them are yours, your share of voice is 15%. It is the comparative metric executives ask for because it reads like market share and moves when a competitor gains ground even if your own mention count holds.

Two variants exist in the tools. Raw-mention share counts every mention equally, and impression-weighted share multiplies each prompt by an estimated search volume so a mention on a high-volume prompt counts for more. Pick one, record which, and never compare a raw figure to a weighted one.

What is the difference between mention rate and citation rate?

Mention rate is the percentage of sampled answers that name your brand, with or without a link. Citation rate is the percentage of sampled answers that link to a page on your domain. Both use the same denominator (prompts × engines × runs in the window), so you can read them side by side.

Compute each per engine and per prompt type before you roll up. A single blended rate hides the fact that ChatGPT and AI Overviews behave differently, and that comparison prompts name brands far more often than problem-solving prompts do.

How should you score visibility and prominence?

For your own visibility score, you can use a weighted composite that adds prominence to the yes/no question of whether you appeared. One option weights each mention by position (first brand named scores 1.0 and second 0.5, with third scoring 0.33 and subsequent positions following the same pattern), averages across the response set, and indexes the result against the top competitor.

Keep mention rate and citation rate visible underneath the composite. When a composite drops 8 points, you cannot tell from that number whether you lost mentions or links. Check position too: moving from first to fourth requires a different fix.

What are you actually measuring with mentions vs citations?

Being named and being linked are different events with different benchmarks, and the gap between them is wide enough to flip a verdict. In Semrush’s ghost-citation study, commercial prompts had a 35.6% mention rate and an 84.4% citation rate. The engines were pulling brand pages as sources far more often than they were recommending the brand by name.

The reverse case shows up at scale too. In Writesonic’s analysis of 9 million AI answers, up to 75% of answers that cited a brand’s page never named the brand. You can record that citation even when the answer never names your brand. Track name-only mentions separately from links to your pages.

Choose which lever to pull based on the rate where you trail. Build entity signals through reviews and listicles, alongside third-party coverage that associates your brand with the category. For citations, structure pages so the retrieval layer can extract a 40–60 word chunk. Report both and benchmark both against their own reference values. Never let one stand in for the other.

Once you can read mention rate and citation rate as separate lines, the next job is what AEO actually is, then how to choose the prompts you track, then which AI visibility metrics belong next to traffic.

What are the AI visibility benchmark ranges by industry?

The best vertical citation benchmark currently published comes from SUMAX’s State of AI Search 2026, which ran 12,000 prompts across 150 brands and six engines in two 14-day waves and reports the median citation rate per industry along with the 75th percentile and top decile:

B2B software sat at a 24% median, 41% at the 75th percentile, and 58% in the top decile. E-commerce and retail sat at 22%, 39%, and 56%. Finance sat at 19%, 33%, and 51%.

Read the 75th percentile as good and the top decile as the ceiling a category leader holds. A B2B SaaS brand cited in 41% or more of category answers across those six engines is ahead of three quarters of its peers in that sample.

Mention rate and citation rate by vertical come from a different panel with a different denominator, so keep them on a separate line in your report. Victorious’s Q1 2026 study of 177 brands and 107,011 responses across eight platforms recorded these averages:

• SaaS: 24.3% mention rate and 13.8% citation rate, with the ChatGPT-only mention rate at 27.8%.
• Financial services and retail: Financial services recorded a 20.1% mention rate and 21.9% citation rate, one of the few verticals where links outrun names. E-commerce and retail recorded an 18.1% mention rate and 6.0% citation rate.

Distribution: Victorious researchers also found a skewed distribution: 89.8% of brands had no AI mentions at all, and only 18 of the 177 had a mention rate above zero. Averages in this field are carried by a handful of brands, so “at the vertical average” is already a strong position for a newer company.

B2B SaaS: PartnerStack’s 2026 AEO benchmark tracked 1,259 brands and 4.3M+ responses across eight platforms and found the median brand appeared in 31.2% of responses, with the top 10% appearing in 63.8% or more. The bottom 10% appeared in 5.7%.

Brands starting their first audit averaged a 19.4% mention rate. Per model, Grok appeared most generous at 41.7% and Perplexity least at 21.3%.

These three studies do not reconcile with each other, and nobody has published a conversion between them.

How does visibility differ across ChatGPT, Perplexity, Gemini and AI Overviews?

A brand mention carries a link five times as often on Perplexity as on Google AI Overviews, and the engines barely share sources. In Ahrefs’ analysis of 31,000+ brand mentions, a link accompanied the mention 10.7% of the time in AI Overviews, 16.8% in Gemini, 26.9% in ChatGPT, 36.8% in AI Mode, and 51.6% in Perplexity.

Those are link rates for brand mentions rather than engine-wide citation rates, but they set the expectation: an AI Overviews mention-rate benchmark will sit well above its citation-rate benchmark, and a Perplexity program should track the two closer together.

Gemini and ChatGPT lean in opposite directions. The same ghost-citation study found that Gemini named brands in 83.7% of appearances but linked them in 21.4%, while ChatGPT linked in 87% of appearances and named the brand in only 20.7%.

Source overlap between engines is low enough that a one-engine baseline tells you little about the others. Wellows’s study of 22.7M citations from January to June 2026 measured website overlap pairwise:

• AI Overviews and AI Mode shared 24.5% of cited websites, the highest pair.
• AI Overviews and ChatGPT shared 6.9%.
• ChatGPT and Perplexity shared 6.5%.
• Gemini and Perplexity shared 7.4%.

Across all five engines in that dataset, 10.9% of websites and 30.3% of named brands were shared. Track all four engines in scope from day one, or your competitive map has three blind quadrants.

Cross-engine evidence is still thin. No published study runs all four engines on a single prompt set with disclosed sample, dates, and model versions, so each figure above comes from a different corpus and reads as a direction, never as a four-way league table.

Why are traditional SEO rankings not a benchmark for AI search?

Rank is a deterministic position in a fixed list, and an AI answer is a probabilistic draw from a retrieval set that changes between runs. The same prompt returns a different brand list on consecutive runs often enough that a single observation is noise, which is why position 3 means something in a SERP and nothing in a ChatGPT answer until you have averaged many runs. CTR fails for the same reason: there is no stable blue link to click, and a large share of answers resolve with no click at all.

Researchers found limited overlap between organic rankings and AI citations, including on Google’s own engine, where overlap should be highest. For brand-level outcomes the relationship is weaker still: in the Victorious Q1 panel, domain authority score correlated at 0.108 with mention rate and 0.017 with citation rate.

Keep organic rank in the SEO section of the QBR and give AI visibility its own section with its own benchmark. When you report only rank, you miss the citation gap for a rank-1 page that AI engines never cite. You also undervalue a page that ranks 40th but gets pulled into AI Overviews on every run.

How do you set your own baseline with prompts, sampling and cadence?

Prompt tracking is an imperfect science. The same prompt can name a different set of brands on the next run, vendors count mentions and citations differently, and no two published studies share one prompt set. You still run a fixed library so a week-to-week move is a real change in your sample, not a new question list. Treat the output as a range, not a rank. A working spec is 25–50 prompts, all four engines, five runs per prompt per engine, weekly collection, and a 30-day window before you put a number in a QBR. Kevin Indig’s prompt-tracking method lands in that range. His B2B SaaS example uses 40 prompts: 12 brand, 12 category, and 16 problem prompts, each run five times per platform every week. Use it as a starting mix, not as a standard the engines obey.

Keep the library in a shared document with the scoring rules. The tool you use to run it will change. The value is a stable sample, not a precise instrument.

Prompt types: Build the library from buyer prompt types rather than keywords. Use commercial category prompts (“best X for Y”) and comparison prompts (“A vs B”) to measure share of voice. Use informational problem prompts (“how do I…”) to measure citation rate, because the engine needs a source to answer them.

Brand weighting: Keep brand prompts at a minority weight. Peec AI’s prompt-tracking analysis reports that its researchers found 40% of non-branded prompts are contested, meaning the brand appears in 10–90% of runs, against 10% of branded prompts. With a branded-heavy library, you get a flattering, stable number that stays flat when nothing in your market has changed.

Hold brand prompts to a minority of the library so the score cannot hide behind questions that already name you. Indig’s 12 of 40 is one way to do that. Branded prompts measure recognition. Unbranded prompts measure competitive position. Do not average them into one headline number.

Spend budget on prompt breadth before run depth. In the same prompt-tracking analysis, researchers calculated ±6.1 percentage points of monthly uncertainty for 50 prompts run daily, against ±13.2 points for 10 prompts run five times a day. They also calculated that pushing a single 50%-visibility prompt to ±10 points takes about 100 runs.

Add prompts until the portfolio error is tolerable, then add runs only for the handful of bottom-of-funnel prompts where a per-prompt number matters to the pipeline conversation.

A filled-in baseline spec for a B2B SaaS team looks like this:

• Prompt library: 40 prompts. Include 12 brand and 12 category prompts, with the remaining 16 focused on problems. Store them in one shared document with the engine list and scoring rules.
• Engines: ChatGPT, Perplexity, Gemini, and Google AI Overviews, logged from a clean, memory-off browser profile when run manually.
• Sampling: Five runs per prompt per engine, collected on the same weekday, for 800 answers a week.
• Metrics: Mention rate and citation rate, plus share of voice against five named competitors. Break each out by engine and prompt type.
• Composite: Weight mention rate at 50% and citation rate at 30%. Give position-weighted prominence the remaining 20%, then index to the top competitor at 100.
• Volatility: The range of the weekly mention rate across the four weeks of the window, reported next to the mean.
• Baseline window: 30 days, with the first reported number being the four-week average rather than any single week.

Those composite weights are a deliberate choice you should make and document, since no published study validates one weighting over another. The volatility line is what keeps the number credible in a QBR: if share of voice jumps in a single week with no publishing or PR change on your side, read it as measurement variance until the four-week average moves.

Manual tracking covers the first month. Use a spreadsheet to log the 800 answers a week in this spec and spend the month learning which prompts are contested. Once the library stabilizes, move to a tool, keep the same prompts.

Which tools measure AI visibility benchmarks?

The tools covered here are Semrush’s AI Visibility Toolkit (Semrush is now an Adobe company), Ahrefs Brand Radar, Profound, and Peec AI. The differences that matter for benchmarking sit in the denominators rather than the dashboards, and each vendor’s documentation describes its own math differently.

• Semrush AI Visibility Toolkit: Defines mention rate as the percentage of AI answers to tracked prompts that mention the brand and share of voice as the percentage of mentions versus competitors, with daily prompt tracking and weekly brand performance reporting.
• Ahrefs Brand Radar: Separates “Found in” (brand mentioned) from “Cited in” (domain linked) and weights AI share of voice by estimated impressions derived from Google search volume, so high-volume prompts count more.
• Profound: Computes visibility score as responses mentioning the brand divided by responses mentioning at least one brand, so answers with no brands at all drop out of the denominator, and runs every tracked prompt once a day.
• Peec AI: Treats one prompt run on one engine on one date as the unit, computes share of voice as your mention count over the sum of all tracked brands’ mention counts, and reports a simple average position when mentioned.

None of these numbers can be compared to a benchmark from a different tool without a translation step. When you pick one, write its definitions into the same document that holds your prompt library, so the people reading your QBR eighteen months from now know what the series measured.

What actually improves your benchmark?

Build third-party presence to improve mention rate and structure pages for extraction to improve citation rate. Researchers have demonstrated no clear effect of schema markup on either. The evidence for each lever is uneven, so the sequence below runs from strongest to weakest.

Third-party validation is the mention-rate lever for B2B. In the Writesonic dataset cited earlier, 82% of bottom-funnel commercial citations went to third parties and owned pages received 3%. Citation lift by presence on each platform in that study:

• YouTube presence corresponded to a 2.8× lift.
• LinkedIn presence corresponded to 2.7×.
• Reddit presence corresponded to 2.4×.
• G2 presence corresponded to 1.9×.
• Gartner presence corresponded to 1.6×.

Content chunking is the citation-rate lever, and the mechanism is retrieval. Engines pull 40–60 word chunks and synthesize across sources, so a page gets cited when one of its sections answers the fan-out query on its own. In AirOps’ fan-out analysis of 50,553 ChatGPT responses, pages whose primary heading matched the retrieval query at 0.90+ similarity had a 41.0% citation rate, and 500–999 word pages were cited at 34.3% against 28.6% for pages over 5,000 words. AirOps researchers also found that word count did not separate always-cited from never-cited pages. Neither did headings or lists, which means heading-to-query match is the variable to engineer, not length.

Structured data: Keep schema in the technical-accessibility column rather than reporting it as a visibility driver. Ahrefs’ matched-control test compared 1,885 pages that added JSON-LD against 4,000 matched controls over 30 days and found AI Mode at +2.4% and ChatGPT at +2.2%, both indistinguishable from zero, and AI Overviews at −4.6% relative to controls.

A separate 6-million-URL cut found cited pages almost 3× more likely to carry JSON-LD, which is an association, not an effect. Keep schema for crawlability and stop reporting it as a citation driver.

Outcomes: Benchmarks for B2B are mixed enough that you should report conversion rate, never volume. In SEO Works’ 95-site referral benchmark, AI referrals converted at 7.18% against 6.08% for organic pooled, but B2B products came in at 0.62× organic (AI converted worse on 6 of 10 sites) while B2B services came in at 1.21×.

No study has published a per-engine click-through rate on AI citations with a disclosed method, so leave CTR off the scorecard until one exists.

How do you start an AI visibility benchmark?

Write the prompt library before you open any tool. Forty prompts in Indig’s 12/12/16 split is a workable first pass, not a certified set. Add five competitors, the four engines, and the scoring rules to one shared document your content, SEO, and product marketing leads can all read. Everything downstream, a spreadsheet or a vendor dashboard, is only a way of executing that sample.

Pick the one published benchmark whose denominator matches your setup and name it in the report header. A six-engine citation program compares against the SUMAX B2B software row, while an eight-platform appearance program compares against the PartnerStack median. A mention-rate program compares against the Victorious SaaS figure. Run the library manually for 30 days and report the four-week average per engine with its volatility range. File it in the QBR as its own channel line beside organic. The weekly operator notes on how teams actually run that scorecard land in The Messy Middle newsletter.

After the first baseline, read the two rates separately and commit the next quarter to whichever is further behind its benchmark. For a mention gap, work on G2 and YouTube presence. Add listicle placements alongside that work. For a citation gap, rewrite your own pages’ headings to match retrieval queries, one section at a time. Rerun the baseline library and compare the rates four weeks later to assess the result.