AI Citation Rate Benchmarks by Brand Size 2026: Startup vs Growth vs Enterprise (200-Site Cohort Study)

Part of the generative engine optimization guide.

I have written four or five articles in the past six months that all dodged the same operator question because I did not have a clean dataset to answer it. The question is some variation of: "We are a 12-person SaaS startup. Our ChatGPT citation rate is 2.3%. Is that good?"

Honest answer for most of 2025: I have no idea. I had per-engine medians from our vertical-cut study, and I had revenue data from the traffic and revenue benchmark, but neither answered the size question. Peec.ai's 1-million-citation benchmark study gave a headline median that conflated sub-10 founders with 5,000-person SaaS giants, and Profound's public index numbers did much the same. Operators kept asking the size question. The published benchmarks kept averaging the answer away.

So between January and April 2026 we built the dataset that answers it. The roughly 200 Attrifast-instrumented sites in our cohort got classified by company headcount, ARR band, Ahrefs Domain Rating, and content age, and we cut citation rate against each axis. This article is the report. The headline finding is simple and stays consistent across every cut: brand size is the single largest determinant of AI citation rate, and the public headline benchmarks are right on average and wrong for most individual operators, because they aggregate across structurally different ceilings.

If you are reading this to figure out whether your number is good, jump to the section that matches your size band. If you are reading it to understand why the published headlines mislead, the methodology and the cross-source reconciliation section are where the real argument lives. And if you are reading it to figure out what to do with the number once you have benchmarked it, the decision tree near the end gives the playbook by band.

Headline numbers

Before any methodology, here are the four numbers I want you to leave with. Everything else in this article is the supporting evidence for these.

The interpretation is the headline finding compressed into one row at a time. A sub-10 founder asking "is 2% good" is asking against the third row of that table when the relevant row is the first. The cohort median of 4.1% is a real number, but it is a center of mass that does not match the experience of any individual operator. The point of this study is to give every operator the right row.

What "AI citation rate" means in this study

Definitions first because the whole study is comparative. Citation rate, for our purposes, is the share of buyer-intent prompts in a tracked prompt set where your domain appears as a citation in the AI engine's answer.

The definition matches what Peec.ai uses in their citation rate benchmark, so the cross-source comparisons later in this article are clean. Profound's index uses a similar definition with minor wording differences. SEOcrawl's prompt tracking and Otterly's reporting both use the same per-prompt presence definition we do. Where methodologies diverge, the differences are usually around prompt selection and engine routing, not the citation-rate metric itself.

A few things citation rate is not. It is not citation share of voice (which is your slice of the cited domains within a single answer). It is not citation density (the median number of unique domains cited per answer, which we covered in the vertical-cut study). It is not click-through rate from a citation. And it is critically not revenue. Citation rate is the first checkpoint in a four-step funnel: presence, click, session, payment. This study measures only step one. The AI visibility KPI guide walks the full stack of metrics if you want to put this number in context.

Methodology

The methodology section is the part most operators skip and most reviewers read first. I have tried to make it specific enough that any team running similar instrumentation could replicate this study against their own customer base and produce comparable numbers.

Cohort construction

The 200 sites in this study are a subset of Attrifast's customer base that consented to anonymized inclusion and that had at least 60 days of citation tracking history at the snapshot moment. We deliberately did not weight by ARR or by Attrifast plan tier because that would have biased the cohort toward larger customers and re-imported the exact aggregation problem we are trying to fix.

Brand-size band definitions

The four bands are deliberately coarse because finer cuts (say, 50-100 versus 100-200) produced cohort cells of fewer than 10 sites where the percentile estimates became unstable. The chosen bands match how operators actually self-describe their company size and how venture and growth-stage capital is typically organized.

The ARR ranges are approximate because we only have self-reported or public-data ARR for about 130 of the 200 sites. Where ARR is known, the correlation between ARR band and headcount band is roughly 0.81, so the two cuts produce similar pictures. We use headcount as the primary axis throughout because it is the more universally available signal.

Prompt corpus

For each site we used a per-site prompt corpus of 80 buyer-intent prompts, selected from the same template families we used in the vertical-cut study. The corpus mix was constructed once per vertical and re-used across sites in that vertical with brand-specific substitutions only on alternatives and comparison templates.

The corpus size of 80 prompts per site times 4 engines times 3 runs gives 960 observations per site, and 192,000 observations across the cohort. That is enough to estimate per-site citation rate to roughly ±0.7 percentage points of true rate, which is the noise floor for any individual rate we publish.

Sources of validation

We checked our cohort medians against three external public datasets, all of which use compatible definitions. The reconciliation comes up later in its own section.

Where this study disagrees with those sources, it is almost always because the public sources blend brand sizes and we do not.

What this study is not

• Not a query-volume-weighted benchmark. Every prompt contributes equally, regardless of underlying real-world query volume.
• Not a measurement of citation quality. Being cited in position 1 versus position 5 of the answer is not measured here; we only count presence.
• Not a revenue study. Citation presence is step one of a four-step funnel; the revenue benchmark handles step four.
• Not a single-engine deep dive. A study designed around ChatGPT alone could go much deeper on conversation context and prompt taxonomy.
• Not a global brand sample. The cohort skews B2B SaaS, fintech, and consumer software because that is where Attrifast's customers sit.

Finding 1: Citation rate scales roughly linearly with headcount

The single largest determinant of AI citation rate across the cohort is company headcount. The relationship is not perfectly linear (the curve flattens at the top), but it dominates every other factor in the dataset including Domain Rating, content age, and engine mix.

The most useful way to read that chart is not the medians, which are obvious, but the overlap zones. The p75 of the solo band (2.1%) is below the p25 of the growth band (1.9%) — almost no overlap. The p75 of the growth band (5.6%) is below the p25 of the mid-market band (3.7%) — substantial overlap. The p75 of the mid-market band (10.4%) overlaps cleanly with the p25 of the enterprise band (9.1%). The bands matter less at the top of the range and more at the bottom: a solo founder genuinely cannot reach the typical mid-market number, but a strong mid-market site can outperform a weak enterprise site.

The mean exceeds the median in every band, which means the distribution has a right tail. The top decile in each band is meaningfully ahead of the median, and the outperformers tend to share traits that are not size-dependent: clean entity disambiguation, real Reddit and Hacker News presence, and a content library that is genuinely best-in-class on a narrow topic. We come back to those traits in the decision tree.

The cohort medians also reconcile against the public benchmarks within tolerable error. The Peec.ai blended median falls between our growth and mid-market band medians (closer to growth), which makes sense because Peec.ai's customer base skews mid-market by absolute count. Profound's index numbers are similar. None of these public sources is wrong, they are just centered on a different point of the distribution than any individual sub-10 operator occupies.

Finding 2: ARR band tells a similar story, slightly noisier

The ARR cut is the cleanest robustness check for the headcount finding. If citation rate is fundamentally a brand-size phenomenon, an entirely different size proxy should produce the same shape, and it does.

The cohort cell sizes are smaller here (130 sites versus 200) because only about two-thirds of customers provided usable ARR data. But the curve has the same shape as the headcount cut: roughly 10x growth from the smallest to the largest band, with a flatter top.

A few things worth flagging from the ARR cut. The $0-100k band sees the highest within-band variance relative to the median, because there is a real difference between a site with $20k ARR that has a niche technical audience and a site with $80k ARR in a generic consumer category. The $10M+ band is the smallest cell and the noisiest, with p25-p90 spanning more than 2x. And the cohort-wide median on the ARR cut (3.9%) is slightly lower than the headcount cut (4.1%) because the known-ARR subset under-represents the largest enterprises.

Finding 3: Domain Rating predicts citation rate, but weakly above DR 60

Domain Rating is the size proxy operators reach for first because it is the easiest to look up. Across the cohort it does carry signal, but less than most assume, and it stops being useful as a marginal predictor above DR 60.

The Pearson correlation between DR and blended citation rate across the 200 sites is roughly 0.46, meaning DR explains about 21% of the variance and the remaining 79% lives in other factors. That correlation is enough to make DR useful as a coarse sort key, but nowhere near deterministic.

Pay attention to the slope. From DR 0-20 to DR 20-40 the median citation rate rises by 1.5 percentage points; from DR 20-40 to DR 40-60 it rises by 2.7 points; from DR 40-60 to DR 60-80 by 3.3 points; and from DR 60-80 to DR 80+ by only 3.3 points despite covering a 20-point DR range that is twice as wide as the others. The slope flattens. Each additional DR point buys less citation rate at the high end than it does in the middle. This matches the Ahrefs DR methodology documentation which explicitly notes that DR is logarithmic and that the gap between DR 70 and DR 80 represents far more link equity than the gap between DR 30 and DR 40.

The implication is uncomfortable for high-DR sites. If you are at DR 75 and pouring resources into a link-building campaign to reach DR 85, the AI citation rate ROI of that campaign is structurally low. The marginal next dollar would buy more citation rate spent on structural content work, brand-mention investment, or vertical-specific editorial PR. The AI citations vs backlinks piece walks the underlying reason: AI engines weight structure, freshness, and entity recognition far more than they weight raw authority once authority crosses a baseline threshold.

Finding 4: Content age compounds citation rate over years

The third axis worth cutting is content age, defined as the median publish date of pages on the domain. This is the noisiest axis because content age and brand size are correlated (older sites tend to be bigger), but the within-band picture is still useful.

The biggest jump is from the under-6-month band to the 6-month-to-2-year band: median doubles from 1.8% to 3.6%. The next jump (to 2-5 years) is roughly 1.5x. Then the curve flattens. The 5+ year band shows a much smaller lift than the previous two, which is consistent with the training-corpus theory: pages need long enough to be ingested by multiple LLM training snapshots, but after a certain point additional age stops adding new training-corpus presence.

Within each content-age band, the within-size distribution still holds. A 5+ year site at solo headcount still tends to underperform a 2-year mid-market site, because the size effect dominates. The cleanest way to read this axis is as a multiplier on the size-band base rate, not as a standalone predictor. A solo site under 6 months old can expect roughly 0.6-1.0% citation rate. A solo site 5+ years old can expect 2-3%. The size band sets the ceiling; content age moves you within that ceiling.

Finding 5: Per-engine variance within a brand is huge

The cohort-wide blended rate hides a per-engine story that matters enormously for any team deciding where to invest. A typical site sits at very different citation rates on each of the four engines, and the variance does not collapse with size.

Two things to read out of that table. First, the per-engine ranking is identical at every size band: Perplexity > ChatGPT > Claude > Gemini, in every row. That stability is the strongest evidence that the engine ranking reflects structural product choices (Perplexity is built as a retrieval-first answer engine, Gemini concentrates citations on a narrower trusted pool), not vertical or brand artifacts. Second, the max/min ratio compresses as size grows. A solo founder might be at 2.3% on Perplexity and 0.6% on Gemini, a 3.8x ratio. An enterprise is at 16.4% on Perplexity and 7.0% on Gemini, a 2.3x ratio. Bigger brands win on Gemini disproportionately because Gemini's trusted-source pool reflects scale and editorial coverage.

The strategic read of that heatmap is engine-specific investment. If you are a growth-stage SaaS at 5.4% on Perplexity and 1.9% on Gemini, the marginal next investment depends entirely on which engine drives your highest-converting traffic, which only a first-party traffic source breakdown can tell you. Some categories see most of their AI conversion traffic from Perplexity; some see it from ChatGPT; very few see meaningful conversion volume from Gemini today, though that share is growing as Gemini integrates more deeply with Google Search.

Finding 6: The cohort distribution is sharply right-tailed

The cohort-wide percentile distribution shows how thin the top of the distribution actually is. Most operators sit well below the headline median, and a small number of outliers carry the average up.

The skew matters because most operators benchmarking against "the median" are benchmarking against a number that more than half the cohort sits below. The mean (5.9% blended across the full cohort) is meaningfully higher than the median (4.1%), which is the classic signature of a right-tailed distribution. If someone reports an "average" citation rate without specifying whether it is the mean or the median, you are missing a 40%+ relative difference between the two.

For operators reading this to set internal targets, the percentile is more useful than the median. A growth-stage SaaS at 5.6% (the growth-band p75) is in the top quarter of its size cohort, even though it is below the cohort-wide median. A solo founder at 2.1% (the solo-band p75) is in the top quarter of theirs, even though that number would be a disaster for an enterprise. The percentile-within-band is the right lens.

Cross-source reconciliation: us versus Peec, Profound, SEOcrawl

This is the section the reviewers will scrutinize. If our 4.1% blended median is way off from what other public benchmarks report, the methodology is in question. The headline answer: it lines up.

The Peec and Profound numbers come in slightly higher than ours, by about 0.4-0.9 percentage points. That delta is almost certainly explained by their cohort mix skewing more mid-market than ours. If we restrict our cohort to the 50 mid-market sites alone, the median jumps to 6.2%, which puts the public benchmarks below our band-specific median, exactly the direction you would expect.

Where this study adds value over the public benchmarks is not in the absolute headline number but in the cross-sectional breakdown. None of Peec, Profound, or SEOcrawl publishes a brand-size cut publicly. The conversation gets stuck at "is 4% good?" until someone publishes the cross-section, which is what this report does. If you are at Peec or Profound and reading this, your underlying data could produce the same cross-section at much larger scale; we would link to it if you publish it.

It is also worth noting where this study is more conservative than the public benchmarks. We use a strict per-prompt-run dedup, we exclude internal-engine self-citations, and we average across three runs rather than reporting a single-shot rate. All three choices push our numbers down slightly compared to less-conservative methodologies. The fact that we still arrive within ±15% of the public headlines is the cleanest possible validation.

What's a good citation rate for you — a decision tree

The whole point of cutting the data this way is to make the "is my number good" question actually answerable. Here is the decision tree, expressed as a table because tables are easier to navigate than nested flowcharts.

Use that table as a sanity check, not as a target. If you are at the high end of the range for your row, you have a real differentiator worth understanding. If you are in the middle, you are on-cohort and the next investment should be the one tied to your weakest engine or your largest revenue source. If you are below the low end, the issue is almost always one of the four levers we have walked: entity disambiguation, structural content quality, brand-mention density, or training-corpus presence.

Why the public benchmarks mislead small operators

If I had to compress the operator-relevant argument of this study into one section, it is this one. The Peec, Profound, and SEOcrawl headline numbers are not wrong. They are technically correct and structurally misleading for any operator whose size does not match the implicit median of the publishing cohort.

The fix is not to discount the public benchmarks. The fix is to read them with the brand-size lens applied. Their data is large and well-collected. Their cuts are useful for the questions they explicitly answer (per-engine variance, per-vertical concentration, source-type mix). Their cuts are not useful for "is my number good for my size," and using them that way produces wrong conclusions on both ends: small operators feeling broken when they are on-cohort, enterprises feeling complacent when their band-specific peers are pulling ahead.

This problem is not unique to AI citation benchmarking. The same issue shows up in SaaS conversion rate benchmarks where founders compare themselves against averages that mix in $1M ARR companies with $1B ARR companies. The remedy is the same: cut by size first, report the cut, and only then collapse to a headline number with the right caveats.

How to actually move your number

The decision tree tells you what range to expect. This section tells you what to do if you are below it.

Two things to flag about that table. First, the effect sizes are additive but not infinitely so. Stacking three or four of these gets you most of the lift; stacking eight does not get you 8 times the lift, because the engines have a ceiling on per-query citation slots. Second, the effect sizes are size-band dependent. A solo founder running the founder-podcast play sees larger relative lift than an enterprise running the same play because the enterprise already has the brand-entity recognition the podcast circuit is building.

The fastest paybacks at the small end of the cohort are the FAQ schema and direct-answer rewrites, both of which are structural rather than relationship-driven. The slowest paybacks but the largest absolute lifts are the brand-mention work (Reddit, podcasts, editorial coverage), because those compound through corpus presence over multiple model retrains.

If your team is choosing where to start, the playbook depends on size. Solo and growth-band sites should start with structural work (schema, direct answers, entity disambiguation) because the relationship-driven work has a long latency and assumes you have already qualified for the candidate pool. Mid-market and enterprise sites should start with brand-mention monitoring at scale, because the structural work is already at diminishing returns and the marginal lift is in maintaining mention density.

Variance, noise, and how confident we are

Three honest caveats about how seriously to take these numbers.

The most honest single statement is that the band medians are stable to within roughly ±1.0 percentage points and the p10/p90 estimates within ±2.0 percentage points. The cross-band ranking is much more stable than the precise values: solo < growth < mid-market < enterprise is essentially certain. The exact gap between bands is the noisier estimate.

How citation rate connects to revenue (briefly)

This study is a presence study. Revenue is a separate concern and we have a separate study for it. But because the operator question almost always ends up at revenue, here is the brief connecting thread.

A citation gets you nothing on its own. It gets you the chance at a click. A click gets you nothing on its own. It gets you a session. A session gets you the chance at a conversion. A conversion gets you revenue. Each step has its own conversion rate, and each step's rate varies enormously by brand and by engine. We measured the full funnel for a different cohort in the revenue benchmark study, which reports per-engine session-to-paid conversion rates that range from roughly 1.4% to 4.1% across our customer base.

The implication is that a higher citation rate is necessary but not sufficient for higher revenue. A solo founder at 3% citation rate with strong on-site conversion might out-earn a growth-stage company at 6% citation rate with weak on-site conversion. The cross-band ranking holds in expectation, but the rank ordering at the individual-brand level is much more about the rest of the funnel than about citation rate. The honest scoreboard is still revenue, which is what Attrifast's first-party Stripe join exists to measure.

Implications for GEO strategy

A few takeaways for operators making investment decisions based on this study.

The most underrated implication is the one about DR. Many teams I talk to still treat Domain Rating as the primary citation-rate lever, partly because it is easy to look up and partly because the SEO industry has spent two decades training operators to think this way. The cohort data is clear: DR is one signal among several, it explains only about a fifth of the variance, and the marginal next DR point above DR 60 buys little citation rate. That money is better spent on the structural and brand-mention work the AI citations vs backlinks piece walks through.

What we would change about this study next time

Three things I would do differently if I were starting this from scratch.

The biggest single methodology improvement would be larger enterprise sample size. The percentile estimates at the top of the cohort are wider than I would like, and a follow-up study with a deliberate enterprise-recruitment push would tighten those numbers. The next planned re-run is scheduled for Q3 2026, where we expect to add roughly 40 sites and re-run the full cohort.

How this study was reviewed

Before publishing we sent the draft methodology and headline numbers to four people for review: a senior data scientist at a former employer, two operator-customers of Attrifast in mid-market and enterprise bands, and one independent SEO researcher familiar with the public benchmark literature. Their pushback shaped three changes: clearer language around the right-tailed distribution caveat in Finding 6, an explicit cross-source reconciliation section against Peec and Profound, and the addition of a "what we would change" section.

None of those reviewers signed off on the conclusions, and none of them are responsible for errors. If you spot a methodology issue or a number that does not reconcile, the contact is on the about page and I will publish corrections inline rather than as a footnote.

FAQ

What is a good AI citation rate in 2026?

There is no single good citation rate, which is the reason this study exists. Across the 200 Attrifast-instrumented sites, the cohort-wide median citation rate is roughly 4.1% across the four major engines, with a p25 of 1.7% and a p75 of 8.9%. But that headline number conflates four very different operator situations. If you are a sub-10-person startup, the honest benchmark is 0.8 to 2.1%; anything above 3% is exceptional. If you are at 10-50 people, the median is around 3.4% and a strong rate is 6-8%. If you are 50-500, the median is 6.2% and 10%+ is achievable. If you are 500+, the median is 11.4% and the leaders sit above 18%.

How do you define AI citation rate?

Citation rate is the share of buyer-intent prompts in a tracked prompt set where your domain appears as a citation in the AI engine's answer, averaged across runs and engines unless we explicitly break by engine. So if you track 100 prompts on ChatGPT and your domain appears in 12 of them, your ChatGPT citation rate is 12%. The blended cohort number averages across ChatGPT Search, Claude with web search, Gemini, and Perplexity. We deduplicate to one domain-appearance per prompt-run regardless of how many pages on your domain were cited.

Why does brand size matter so much for AI citation rate?

Three structural reasons. First, AI engines weight brand-entity recognition in their training corpus, and bigger brands have more co-occurrence in pretraining data, podcasts, news, Wikipedia, and Reddit. Second, bigger brands tend to have higher Domain Rating, more pages indexed, and more pages that already rank on classic search, which gets them into the retrieval candidate pool more often. Third, bigger brands have more editorial coverage on G2, Capterra, Wirecutter, and NerdWallet, which capture roughly 22% of all citation slots. None of these levers move much for a 6-person startup in its first 18 months.

How big is the Attrifast cohort and how was it built?

200 sites total, all using Attrifast's first-party AI traffic and citation tracking between January and April 2026. The distribution is roughly 50 solo and small-team sites (0-10 employees), 80 growth-stage sites (10-50 employees), 50 mid-market sites (50-500 employees), and 20 enterprise sites (500+ employees). The sample skews B2B SaaS, fintech, and consumer software because that is where Attrifast's customer base sits, but it intentionally spans the same 12 verticals we used in our earlier vertical-cut study so the numbers reconcile.

How does this compare to the Peec.ai 1-million-citation benchmark?

Peec.ai's report is excellent and the dataset is much larger than ours, but it cuts the data primarily by engine and aggregates across customer brands without breaking by brand size. Their headline median citation rates fall in roughly the same range as our blended cohort median, which is reassuring as a sanity check. Where we diverge is on the implication: their averages, applied directly by a small operator, set expectations that are achievable only for mid-market and enterprise sites. Our brand-size cut is designed to be read alongside theirs, not against it.

What citation rate should a sub-10-person startup expect?

Based on the 50 sub-10-person sites in the cohort, the median blended citation rate is around 1.4% with a p25 of 0.8% and a p75 of 2.1%. Anything above 3% at this size is genuinely exceptional and almost always reflects either a strong technical-content moat, deep Reddit and Hacker News presence, or a founder whose personal brand drives entity recognition. The most common mistake at this size is benchmarking against the cohort-wide 4.1% median and concluding the strategy is broken. It is not broken, the comparison is wrong.

What citation rate should a 10-50 person growth-stage company expect?

Across the 80 sites in that band, the median is roughly 3.4%, with a p25 of 1.9% and a p75 of 5.6%. The top decile sits around 7.8%. This is the band where brand-mention investment starts to compound visibly, Reddit and podcast presence pays back fastest, and structural GEO levers start producing measurable citation lift over a quarter.

What citation rate should a 50-500 person company expect?

Mid-market sites in our cohort show a median citation rate of 6.2%, a p25 of 3.7%, and a p75 of 10.4%. The top decile reaches roughly 14%. At this size most operators have a critical mass of editorial coverage, real Reddit threads with organic mentions, decent DR (typically 50-70), and a content library old enough that training corpora include meaningful coverage.

What citation rate should a 500+ person enterprise expect?

Across the 20 enterprise sites, the median is roughly 11.4%, p25 9.1%, p75 16.8%. The leaders sit above 18%, and a handful of category-leading brands clear 25% on their own branded query set. At this size the lever is no longer earning the next citation, it is defending the share you have and managing engine-level concentration.

Does Domain Rating predict citation rate?

Loosely. Across the full 200-site cohort, the correlation between Ahrefs DR and blended citation rate is roughly 0.46, which is positive but a long way from deterministic. DR explains around 21% of the variance in citation rate, leaving 79% to other factors. The relationship is also non-linear: each additional DR point buys more citation rate in the 40-60 band than it does above DR 80.

How does content age affect citation rate?

Materially. Sites with most of their content under 6 months old show a median citation rate of 1.8%; 6 months to 2 years sits at 3.6%; 2 to 5 years at 5.4%; 5+ years at 7.2%. The lift is steepest from the under-6-month band to the 2-5 year band, then flattens. The mechanism is mostly training corpus presence.

Is citation rate the same across engines for a given brand?

Almost never. Within the cohort, the average within-brand variance across the four engines is roughly 2.7x between the highest and lowest engine. Perplexity is the easiest engine to earn citations on at every size band; Gemini is the hardest. Anyone publishing a single AI citation rate without specifying the engine is averaging four very different distributions.

How does this study connect to revenue?

It does not, directly. Citation rate measures presence, not click-through, and not revenue. The 2026 AI Traffic Revenue Benchmark is our companion study that joins AI engine sessions to Stripe payments for the revenue side. The two studies are designed to be read together.

Can a small brand outperform the brand-size benchmark for its band?

Yes. Roughly 8% of sub-10-person sites clear 3% blended citation rate, and a smaller subset clear 5%. Three traits dominate the outperformer set: a founder with strong personal brand and podcast presence, a tightly written technical content library that is genuinely best-in-class on a narrow topic, and concentrated authentic Reddit and Hacker News mentions in the right communities.

Related reading from the Attrifast research stack

For hands-on tools, see share of voice in AI search and AI visibility score.