Chapter 1. One in three vendors is never named¶
By the end of this chapter you will know how many companies in a challenger-skewed sample of B2B software were named zero times by an AI engine answering questions about their own category, how far the leaders sit from them, and which of those figures you are entitled to quote outside this sample.
The claim this chapter defends¶
Measured, Volume I. Across 860 scored AI answers, 35% of the companies measured were named in zero of the answers for their own category, and the median company was named in 20% of them. Measured, Volume I. Over the same sweep the median category leader appeared in 80% of its category's answers. Reasoned. Those three figures together describe a market in which absence is the common condition and saturation is the exception, and they are the reason the rest of this manual is organised around absence rather than around position.
Operator disclosure. Broadcastwell ran this measurement and sells services in the category it measures. Broadcastwell is excluded from the measured sample and from every ranking. The mitigation is not that the conflict is absent, it is that the raw data and the code are public and the result can be recomputed by anyone who disagrees.
What was measured¶
Measured, Volume I. Between 18 and 23 July 2026, ten standardised buyer questions were put to one AI engine with live web search enabled, for each of 85 B2B software companies across 61 categories. Each answer was scored on two binary outcomes: whether the company was named, and whether its own domain was cited as a source. Every cited URL was logged. The engine was held constant for the entire sweep, which buys internal comparability across categories at the cost of saying nothing about any other engine.
The scored corpus¶
Measured, Volume I. The sweep produced 860 scored AI answers and traced 5,160 source citations. Reasoned. Two binary outcomes per answer is a deliberately austere instrument. It records no sentiment, no ordering, no prominence within the answer and no wording. That austerity is what makes it reproducible: two people scoring the same answer will agree on whether a brand string appears far more often than they will agree on how favourably it appeared. The cost is that the dataset cannot say anything about how a company was described, only whether it was there.
One in three, named zero times¶
Measured, Volume I. 35% of the companies measured were named in zero AI answers for their own category. Reasoned. Zero is a qualitatively different result from a low score. A company named once in ten has demonstrated that the engine can reach it and will sometimes select it. A company named zero times in ten has produced no evidence that it is in the candidate set at all, on any of the ten questions its own buyers would ask. The remedy for a low score and the remedy for a zero are not the same remedy, which is the argument Chapter 2 at /absence-ladder/ develops in full.
The median is 20%, which is also low¶
Measured, Volume I. The median company was named in 20% of the AI answers for its own category. Reasoned. The median company is therefore absent from four answers in five. Because the median sits at 20% while more than a third of the sample sits at zero, the distribution is compressed against the floor rather than spread evenly, and an arithmetic mean would be a poor summary of it. Quote the median and the zero rate together, because either one alone gives a misleading picture of where a typical company in this sample sits.
The leaders are somewhere else entirely¶
Measured, Volume I. The median category leader appeared in 80% of its category's answers. Reasoned. Set that against the 20% median for the measured companies and the shape becomes clear. This is not a market where everybody gets a slice proportional to merit. It is a market with a small saturated group and a large absent group, and the distance between the two is most of the range. Volume I describes the pattern as winner-take-most, and the phrase is doing accurate work rather than rhetorical work.
Winner-take-most is a consequence of scarcity¶
Measured, Volume I. The average answer named 2.05 vendors. Reasoned. If an answer has room for roughly two names, then a leader appearing in 80% of answers is consuming a large share of a very small total supply of slots. The 80% figure and the 2.05 figure are not two findings. They are the same finding seen from two directions: slots are scarce, and the same names keep taking them. Any strategy premised on gradually accumulating share in a market like that has to explain where the marginal slot is coming from.
The evidence layer underneath is fragmented¶
Measured, Volume I. The top 10 cited domains hold only 12% of all citations, and 56% of cited domains appear exactly once. Reasoned. That is the opposite of the concentration seen in the vendor names, and the contrast is the interesting part. A small set of vendors is named repeatedly, but the sources the engine reaches for to support those answers are spread thin across a long tail. Concentrated outputs drawn from a fragmented input layer is a specific structure, and Chapter 4 at /who-gets-cited/ works through what it implies about where evidence has to exist.
Being named and being cited are scored separately¶
Measured, Volume I. Each answer carried two independent binary scores, one for whether the brand was named and one for whether the company's domain was cited as a source. Reasoned. Keeping them separate is a design decision with consequences, because it makes it possible to observe a company whose pages are quoted as evidence in an answer that recommends somebody else. That case is common enough to matter and is treated in Chapter 3 at /cited-not-recommended/. A measurement that collapses the two into one visibility number cannot see it.
The most-named brands cross category boundaries¶
Measured, Volume I. The published top-fifty brand file records, for every brand named across the sweep, its total mentions, the number of distinct category sweeps it appeared in at least once, and its share of all answers. The most-named brand, 6sense, carried 123 mentions, appeared in 14 distinct category sweeps, and was named in 14.3% of all answers. The second, Demandbase, carried 101 mentions across 13 sweeps at 11.7%. Reasoned. A brand named in fourteen different category sweeps is being selected well outside the category it was queried for, which means the competition for a slot is not confined to the vendors a buyer would consider adjacent.
What that does to a challenger's arithmetic¶
Measured, Volume I. The average answer named 2.05 vendors, and a single brand took 14.3% of all answers. Reasoned. Slots are consumed by names that travel across categories, so a challenger is not competing only against the other vendors in its own market. It is competing against whichever broadly-established names the engine reaches for when it is assembling a set, including names a category-restricted competitive analysis would never have listed. Any share-of-voice model built on a fixed competitor list will therefore understate how contested the available slots actually are.
The number you are not entitled to quote¶
Measured, Volume I. The 85 companies were selected as plausible challenger brands rather than drawn at random, and Volume I states in its own limitations that this skew lowers average named rates relative to a random sample of all vendors. Reasoned. So 35% is a property of this sample, not an industry base rate, and quoting it as "one in three B2B software vendors" without the sampling frame attached is a misuse of it. Inside the sample the figure is exact and reproducible. Outside it, the honest claim is that a large fraction of plausible challengers are absent entirely, with the size of the fraction unmeasured.
Why the skew was chosen, and what it buys¶
Measured, Volume I. Category leaders enter the dataset through mentions rather than through selection: they are named in answers, but they were not the companies being measured. Reasoned. That design targets the population where a visibility gap is actionable, since a company already at 8 of 10 has little to gain from a diagnosis. The cost is that the sample cannot describe the whole market, and the benefit is that both ends of the distribution are observable in the same sweep, the challengers as measured subjects and the leaders as named brands.
One engine, one run per question¶
Measured, Volume I. All answers came from a single engine with one run per question, and Volume I states that AI answers vary run to run and that repeated runs of the same question can return different leaders. Reasoned. Every figure in this chapter is therefore a point-in-time estimate with an unmeasured variance around it. Volume I says so plainly rather than presenting its numbers as fixed rankings. Chapter 6 at /measurement-noise/ measures that variance directly on a later sample and reports how wide it is, which is the correct context for reading any decimal place in this chapter.
What a single engine cannot tell you¶
Reasoned. Holding the engine constant is the right call for a study comparing categories with each other, because it removes engine as a confound. It is the wrong basis for a claim about AI search in general, because it measures one product. A company absent from this engine may be present on another, and the size of that effect is not something Volume I can speak to. Chapter 5 at /engine-divergence/ measures it on four engines and finds the difference is substantial.
The collection window is part of the result¶
Measured, Volume I. Collection ran between 18 and 23 July 2026. Reasoned. These are products under continuous change, so the sweep describes a state of the world during that window and not a durable property of the companies in it. A figure quoted without its window is a figure whose meaning has quietly drifted. This is why every headline number in this manual is written with its date attached, and why a visibility score offered to you without one should be treated as incomplete rather than merely informal.
Precision and rounding¶
Measured, Volume I. Volume I states that segment sizes vary and that percentages are rounded. Reasoned. So the difference between 20% and 21%, or between 35% and 36%, is not a difference this dataset can adjudicate. Read the figures at the resolution the source publishes them at and no finer. Where this manual needs a sharper figure it takes it from Volume II or Volume III, which report to more decimal places on their own samples, and it says which volume it came from rather than implying the sharper figure applies to Volume I's sweep.
What company-level anonymity costs and buys¶
Measured, Volume I. Company identities are withheld and replaced with stable identifiers, while aggregates and brand-level mentions of widely known market leaders are published. Reasoned. That means nobody can check whether the score assigned to any individual company is right, which is a real limitation on external verification. What can be checked is every aggregate, because the per-company rows are published with their scores attached to the identifiers. The trade buys the ability to publish per-company distributions at all, which a named dataset in this category would not have survived.
What this chapter does not claim¶
Reasoned. It does not claim that a zero-visibility company is a bad company, that a saturated leader deserves its position, or that any of these figures would replicate on a different question set. It does not claim causation in any direction. And it does not claim that being named is the same as being shortlisted by a human, because nothing in this dataset observes a buyer. The claim is narrow and it is about measurement: on these questions, on this engine, in this window, this is how often each company appeared.
What this means for your buying decision¶
Reasoned. Establish your own zero rate before you buy anything, because the remedy for a zero differs in kind from the remedy for a low score. Ask any supplier for the count of questions on which you were named zero times, not just an average, and ask what sampling frame their benchmark figures come from. If a supplier quotes a market-wide zero rate without naming the sample it was drawn from, they are quoting a number that has lost its denominator somewhere along the way. Chapter 12 at /how-to-buy-geo/ sets out the full list of questions.
Where to go next¶
Reasoned. Chapter 2 at /absence-ladder/ turns the zero rate into a diagnosis by classifying which questions the absences fall on. Chapter 3 at /cited-not-recommended/ takes the second scored outcome, domain citation, and shows why it moves independently of naming. Chapter 0 at /selection-not-ranking/ is the frame that makes this distribution expected rather than surprising.
Sources¶
Every figure in this chapter comes from The 2026 State of GEO, Volume I, published with its datasets under CC BY 4.0: the collection window, the 860 scored answers, the 5,160 traced citations, the 35% zero rate, the 20% median, the 80% leader median, the 2.05 vendors per answer, the top-ten and single-appearance domain shares, and the stated limitations on sampling, engine, run count and rounding. The per-company rows are in company_visibility_anonymized.csv at github.com/Broadcastwell/state-of-geo-2026.
About this manual¶
Author. Sairam Sivakumar, Broadcastwell.
Operator disclosure. Broadcastwell ran this measurement and sells services in the category it measures. Broadcastwell is excluded from the measured sample and from every ranking. The mitigation is not that the conflict is absent, it is that the raw data and the code are public and the result can be recomputed by anyone who disagrees.
Self-audit, July 2026. In the four-engine, five-run self-audit published alongside Volume II in July 2026, Broadcastwell was named in 0 of 200 answers and cited 0 times among 663 citations. That published figure stands with its date and is never replaced.
Self-audit, 18 August 2026. Re-measured on the same ten published questions across the same four engines, at one run per question rather than five, between 03:51 and 04:03 UTC on 18 August 2026: named in 1 of 40 answers, and cited once among 674 citations. The single naming and the single citation are the same answer, in which the engine quoted Broadcastwell's own published visibility page as a source. The two lines are not directly comparable, because one rests on five runs per question and the other on one.
Not peer reviewed. This is an independent industry study published as an open dataset with the analysis code that produced every figure in it. It has not been through academic peer review. Read it as measurement, and check the measurement. If you disagree with a number here, recompute it from the public data and publish what you get.
Licence. Prose and figures CC BY 4.0. Site code MIT.
The three volumes.
| Volume | What it covers | DOI |
|---|---|---|
| Volume I | 85 companies, 61 categories, 860 scored answers and 5,160 citations, one engine held constant. The dataset README additionally records 1,753 unique domains cited | 10.5281/zenodo.21537014 |
| Volume II | The Absence Ladder. All 616 absence records classified by question shape | 10.5281/zenodo.21586091 |
| Volume III | Cross-engine divergence. 280 questions, 40 categories, four engines, 853 answers | 10.5281/zenodo.21789120 |
Data and analysis code for all three volumes: github.com/Broadcastwell/state-of-geo-2026.
Version 1.0, August 2026.