Research · Volume II

The Absence Ladder

Name: The Absence Ladder: The 2026 State of Generative Engine Optimization, Volume II
Creator: Broadcastwell
License: https://creativecommons.org/licenses/by/4.0/

How the shape of a company's AI invisibility changes as its visibility rises

Broadcastwell · July 2026

Abstract

Most measurement of AI search visibility reports a single number: how often an engine names a company when buyers ask about its category. That number is useful but flat. It says nothing about which questions a company loses, and therefore nothing about what to do next.

This study takes the inverse view. Across 85 B2B software companies in 60 categories, we recorded every buyer question the engine answered without naming the company, then classified all 616 of those absences by question shape. The result is a systematic and statistically significant pattern we call the absence ladder.

Companies with no visibility lose category-level questions: best-of shortlists and alternatives-to-incumbent queries. Companies with high visibility lose almost nothing except head-to-head comparisons against a named rival. The mix shifts monotonically across every tier in between.

We also report one strong null result. The strength of a category's leader does not predict how invisible its challenger is. AI visibility, in this dataset, is not zero-sum.

B2B software companies

product categories

absence records classified

buyer questions per company

60 categories here versus 61 in Volume I: Volume I's count includes one category that was queued but not scored, and is excluded from this analysis.

Read this first

Four constraints shape everything below. We put them here rather than at the end because they change how the numbers should be read.

·One engine, one run per question. Every answer came from a single AI engine with live web search, queried once. We did not measure run-to-run variance. Treat the direction of each finding as the result, not the decimal.
·Challenger-skewed sample. Companies were selected as plausible non-leaders, so the zero-visibility rate is a property of this sample, not of B2B software generally.
·Small top tier. The highest visibility tier rests on 8 companies and 16 absence records. We also report every test with that tier excluded.
·Classification is rule-based. Question shapes were assigned by regular expression, not human annotation. The rules and their precedence are published below so anyone can reclassify.

Full detail in section 07.

Why absence, not presence

A single visibility score hides the only thing that is actionable

A visibility score of 2 out of 10 tells you that eight answers went by without you. It does not tell you what those eight answers were about. Two companies can both score 2 and be in entirely different strategic positions: one losing every "best X software" list, the other losing only "X vs Y" comparisons against a specific rival.

Those are different problems with different fixes, and a single percentage hides the difference completely.

So we inverted the unit of analysis. Instead of studying the answers where a company appeared, we studied the answers where it did not, and asked what those questions had in common.

Method

How the absence records were produced

Sample.

85 B2B software companies across 60 distinct product categories. The sample is challenger-skewed by construction: companies were selected as plausible non-leaders in categories with an identifiable incumbent, which is the population where visibility gaps are actionable.

Note on the category count.

Volume I reported 61 categories. That figure counts every category string in the collection sheet, including one category whose company was queued but never scored. Volume II analyses only companies with complete scores, which is 60 categories. The 85 companies are the same in both volumes.

Procedure.

For each company we generated ten buyer questions reflecting how a real purchaser researches that category, then ran each question through one AI engine with live web search enabled. For every answer we recorded whether the company's brand was named, whether its domain was cited as a source, and which competitor was named most often and how many times.

Absence records.

For every question where the company was not named, we retained the full question text. That produced 616 absence records. We validated completeness: for all 85 companies, absence count equals ten minus the visibility score, with zero mismatches.

Classification.

Each absence question was assigned one shape by regular-expression rules applied in fixed precedence.

Table A · Classification rules

Data table: Shape, Rule, Example
Shape	Rule	Example
Head-to-head comparison	contains "vs", "versus", or "comparison"	"Whitespace vs Artificial Labs: which is better?"
Alternatives-to-incumbent	contains "alternative"	"What are the best Sequel alternatives?"
Best-of shortlist	contains "best" or "top N"	"What is the best resource management software?"
Evaluation criteria	evaluate, choose, select, criteria, pricing model	"How should I compare pricing models when switching?"
Use case / problem	opens "how can/do/does" or contains "use case"	"How can this software reduce manual data entry?"
Definitional	opens "what is/are" without "best"	"What is reinsurance placement software?"

Observed classification proportion; base: 616 absence records. Interval not reported in the published source.

2.8 percent of records fell to "other". Precedence matters and is stated so the classification is reproducible: a question containing both "best" and "vs" is counted as a comparison.

precedence, applied in this order
1 comparison   vs | versus | comparison
2 alternatives   alternative
3 best-of   best | top N
4 evaluation   evaluate | choose | select | criteria | pricing model
5 use case   how can/do/does | use case
6 definitional   what is/are, without best

Engine.

One engine, one run per question. See section 07.

Finding 1

The absence ladder: what you lose depends on where you stand

The distribution of absence shapes shifts systematically with visibility.

Absent
Named
Recommended
Cited

Four positions, recorded separately. A named answer is not a cited answer.

Figure 1. Line chart. The category-level share of absences falls from 55 percent to 12 percent as visibility rises, while the head-to-head comparison share climbs from 20 percent to 69 percent. — Figure 1. As visibility rises, absences shift from category-level questions to head-to-head comparisons. n = 616 absence records across 85 companies.Observed shares of classified absences. Tier bases: 300, 199, 101 and 16 absences, from 30, 25, 21 and 8 companies respectively. Interval not reported in the published source.

Table B · The absence ladder

Data table: Visibility, Companies, Absences, Category-level, Head-to-head
Visibility	Companies	Absences	Category-level	Head-to-head
Named 0 of 10	30	300	55.3%	20.0%
Named 1 to 3	25	199	48.2%	24.6%
Named 4 to 6	21	101	20.8%	41.6%
Named 7 to 10	8	16	12.5%	68.8%

Observed shares of classified absences. The Absences and Companies columns give each tier base; percentages use that row’s absence count. Interval not reported in the published source.

Category-level = best-of shortlists plus alternatives-to-incumbent queries.

Nine companies scored 7 or above. One of them scored 10 of 10 and therefore produced no absence records at all, so the top tier's absence analysis rests on the remaining 8 companies and their 16 records. Company counts across the four tiers sum to 84 in this table and to 85 in the sample.

Head-to-head share of absences, by tier

Named 0 of 1020.0%

Observed head-to-head share: base: 300 classified absences from 30 companies. Interval not reported in the published source.

Named 1 to 324.6%

Observed head-to-head share: base: 199 classified absences from 25 companies. Interval not reported in the published source.

Named 4 to 641.6%

Observed head-to-head share: base: 101 classified absences from 21 companies. Interval not reported in the published source.

Named 7 to 1068.8%

Observed head-to-head share: 11 of 16 classified absences from 8 companies. Interval not reported in the published source.

The relationship is significant and holds under multiple tests:

·Chi-square across tier by shape: chi-square = 61.8, df = 15, p = 1.3 x 10^-7, Cramer's V = 0.18
·Visibility vs absence-is-category-level: r = -0.284, p = 6.8 x 10^-13, n = 616
·Visibility vs absence-is-comparison: r = 0.234, p = 4.1 x 10^-9
·Excluding the small top tier entirely, the comparison effect survives: r = 0.184, p = 5.9 x 10^-6, n = 600

Figure 2. Stacked bar chart of the question-shape mix for each visibility tier, with segments for best-of, alternatives, comparison, use-case, evaluation and other. — Figure 2. Full question-shape mix for each visibility tier.Observed shares of classified absences. Tier bases: 300, 199, 101 and 16 absences, from 30, 25, 21 and 8 companies respectively. Interval not reported in the published source.

Interpretation

Invisibility is not one condition. It has stages.

A company named zero times is failing at the category door. The engine does not consider it a member of the set when a buyer asks who the players are. More than half its losses are questions that never mention a competitor by name. They simply ask who exists, and the answer does not include it.

A company named seven or more times has cleared that door. It is a recognised member of the category. What it now loses is the second gate: direct comparison against a specific named rival. Nearly seven in ten of its remaining absences are "X vs Y" questions.

The middle tiers sit exactly where you would expect if the two gates are sequential rather than simultaneous.

Practical consequence

The work implied by each stage is different. A company at the category door needs entity presence and inclusion in the third-party sources engines consult when assembling a set. A company at the comparison gate needs comparison content that a model can quote against a specific named competitor. Prescribing the second to a company stuck at the first is a common and expensive mistake, and a single visibility percentage cannot tell you which one you are.

Finding 2

Being cited is not being recommended

Brand naming and domain citation correlate, but loosely: Pearson r = 0.512, Spearman rho = 0.577, p less than 10^-6. They are not the same signal.

Figure 3. Scatter plot of times named against times cited, with a red cluster of companies on the left edge and a dashed diagonal marked named equals cited. — Figure 3. Nine companies were cited as a source while never being named as a recommendation.

of companies were cited more often than they were named

Observed company proportion: 25 of 85 companies. Interval not reported in the published source.

among those, the average company was named 2.2 times but cited 4.4 times

companies were named zero times while their own domain was cited as a source, one of them five times out of ten

Those nine companies were named zero times while their own domains appeared among the cited sources. Source inclusion and vendor naming are separate observed outcomes; these records do not reveal the engine's internal reading or trust decisions.

Practical consequence

Content volume is not the binding constraint for this group. Being quotable and being recommendable are separate properties, and a content strategy aimed only at the first will not deliver the second.

Finding 3 · null result

No significant association detected in this sample

We tested whether leader and challenger visibility were associated. No statistically significant association was detected in this sample; that does not establish that leader strength can never affect a challenger.

Figure 4. Scatter plot of challenger visibility against category leader visibility with an orange mean line and no visible trend. — Figure 4. Leader strength does not predict challenger visibility. Spearman rho = -0.05, p = 0.64, n = 85.

·Spearman rho between leader visibility and challenger visibility: -0.051, p = 0.64, n = 85
·Challenger mean when the leader scores 9 to 10: 2.35
·Challenger mean when the leader scores 7 to 8: 3.48
·Challenger mean when the leader scores 6 or below: 1.57
·Kruskal-Wallis across those bands: p = 0.052

The relationship is not monotonic and is not significant at conventional thresholds. Challengers in categories with a moderately strong leader scored highest on average, and challengers in categories with a weak leader scored lowest.

Interpretation, stated cautiously

In this dataset, an answer naming a strong incumbent does not appear to be an answer that had no room for anyone else. The plausible reading is that a well-defined category with an obvious leader is also a category the engine understands well enough to name a second and third option, whereas a category with no clear leader is often a category the engine does not model cleanly at all. In seven of the 85 company sweeps no brand scored above 5 of 10, and challengers in those categories did worst.

We report this as a null result rather than a finding. It should be retested on a larger sample before anyone builds strategy on it.

Category concentration

How wide the gap between leader and challenger actually is

Figure 5. Bar chart of the distribution of challenger visibility from 0 to 10, with the zero bar highlighted in red. — Figure 5. Distribution of challenger visibility. 35% were named zero times.Observed company proportion: 30 of 85 companies. Interval not reported in the published source.

·Median challenger visibility: 2 of 10. Mean: 2.75.
·35.3 percent of challengers were named zero times in their own category.Observed company proportion: 30 of 85 companies. Interval not reported in the published source.
·Median category leader: 8 of 10. Mean: 7.73.
·In 58.8 percent of the 85 company sweeps the category leader scored 8 or above; in 36.5 percent, 9 or above.Observed company-sweep proportions: 50 of 85 scored 8 or above; 31 of 85 scored 9 or above. Interval not reported in the published source.
·In only 8.2 percent of sweeps did no brand score above 5 of 10.Observed company-sweep proportion: 7 of 85. Interval not reported in the published source.
·The median leader-minus-challenger gap was 5 points; in 23.5 percent of sweeps the gap was 8 points or more.Observed company-sweep proportion: 20 of 85 had a gap of 8 points or more. Interval not reported in the published source.
·In 16 sweeps the challenger scored zero while the leader scored 8 or above.

Limitations

What this study cannot tell you

These matter and we state them plainly.

One engine, one run.

All answers came from a single AI engine with live web search, one run per question. AI answers vary between runs; we did not measure that variance, and we cannot claim these proportions would replicate exactly. Treat the direction of each finding as the result, not the decimal.

Challenger-skewed sample.

Companies were selected as plausible non-leaders. The 35 percent zero-visibility rate is a property of this sample, not of B2B software generally.Observed company proportion: 30 of 85 companies. Interval not reported in the published source.

Small top tier.

Nine companies scored 7 or above, but one scored a perfect 10 of 10 and contributed no absence records, so the top tier rests on 8 companies and 16 absence records. The 68.8 percent comparison figure rests on a small base. The underlying trend survives when that tier is excluded, which is why we report both.Observed head-to-head share: 11 of 16 classified absences from 8 companies. Interval not reported in the published source.

Classification is rule-based.

Question shapes were assigned by regular expression, not human annotation. Rules and precedence are published so anyone can reclassify and check.

Question generation is a design choice.

The ten questions per category were generated to reflect realistic buyer research. A different question set would produce a different absence mix. The shapes are drawn from that set, which is why the tier comparison is internally consistent even if absolute proportions are set-dependent.

Scope.

60 categories is enough to see a pattern, not enough to call any individual category.

Data availability

Every number in this paper is reproducible

Three datasets are published with this paper. Company names are replaced with stable identifiers; question text and all scores are unmodified.

challenger_visibility_v2.csv

85 rows

Per-company: category, times named, times cited, leader score, gap, absence count.

Download CSV

absence_questions_classified.csv

616 rows

Every absence question, its verbatim text, assigned shape, and the company's visibility tier.

Download CSV

absence_shape_by_tier.csv

4 rows

The aggregate table behind Finding 1.

Download CSV

Anyone can reproduce every number in this paper from these three files.

Archived and citable: 10.5281/zenodo.21586091

Published open access under CC BY 4.0 with a DOI. This paper has not been peer reviewed. The classified dataset is public so anyone can rerun the analysis.

Browse the full repository on GitHub

What this changes

AI visibility is not one metric and should not be managed as one

If the absence ladder holds, "AI visibility" is not one metric and should not be managed as one.

The useful question is not what percentage of answers name us. It is what kind of question are we currently losing. That answer tells you whether you are fighting for admission to a category or for preference within one, and those require different work.

The single number is a symptom. The shape of the absence is the diagnosis.

Our current measurement protocol is on the methodology page.

Find the questions your buyers are asking.

See which buyer questions leave you out, the sources visible in those answers, and the three changes worth testing first. The Category Audit covers ten questions; the Diagnostic extends to 35 questions and 525 scheduled observed answers. AI answers vary, so repeated observations and disagreements are reported.

Start with the $490 Category Audit

Credits in full against the Diagnostic within 30 days.

Get the Diagnostic, $990

Credits in full against the program.

Every figure comes from the published method, v1.1.

Free 10-question check

The 2026 State of GEO, Volume II. Broadcastwell, registered in Indiana, United States. Data collected July 2026. Volume I, covering 860 scored answers and 5,160 citations, is published separately with a DOI.