The state of citability on the web
Every month we publish what Primafonte finds across the websites it analyses: how many are ready to be cited by a model, how many are not, and exactly what fails most.
These are aggregate figures from real scans. No individual website or score appears: what is published describes the whole, not anyone in particular.
Sample
36 websites analysed in October 2026, measured with methodology 0.16.0 · provisional thresholds.
Figures are grouped by methodology version. The criteria have changed over time, so mixing versions would make month-to-month comparison meaningless.
How are the scores distributed?
Scores are not grouped by tens but by what they mean: whether a model can use that website as a source, whether it finds it but cannot cite it reliably, whether it works with specific blind spots, or whether it has no structural blockers. The band matters more than the exact number, because it is what says whether there is anything to fix.
- Ready — no structural blockers
- 22 · 61%
- Citable with gaps — it works, with specific blind spots
- 9 · 25%
- Partially readable — they find you, but cannot cite you reliably
- 5 · 14%
- Not citable — models cannot use this site as a source
- 0 · 0%
What fails most
Share of websites that do not pass each check, calculated only over those that could be measured. Inconclusive results are excluded, as they are in the score.
- Links to external sources
- 94% · 33/35
- Question-shaped headings
- 94% · 33/35
- Markdown content negotiation
- 92% · 33/36
- Direct answers after each heading
- 89% · 31/35
- Data in the content
- 86% · 30/35
- Declared authorship
- 83% · 29/35
- Lists and tables
- 69% · 24/35
- A declared stance towards the crawlers that cite you
- 58% · 21/36
- Identifiable company or author page
- 46% · 16/35
- Link discovery headers
- 37% · 13/35
What does almost everyone get right?
Most websites pass these five, and seeing them matters as much as seeing the failures: a check everyone passes distinguishes nobody, and if one sits there long enough we have to decide whether it still contributes or is merely holding weight that another would use better.
- The crawlers that cite you are not blocked
- 100% · 36/36
- Clean redirect chain
- 100% · 35/35
- Same content for bots and for people
- 100% · 35/35
- Paragraphs that stand outside their context
- 97% · 34/35
- The crawler is not blocked
- 97% · 35/36
Which other months have data?
Every month is published separately and with the methodology version it was measured under. Versions are never mixed: the criteria have changed several times, and joining months measured under different rules would give a series that means nothing however tidy it looks.
Where do these figures come from?
The numbers on this page come from real scans run by the engine documented in the methodology. The standards and studies those checks are built on are these, and anyone can check them independently.
- GEO: Generative Engine Optimization
The study that measured the effect of citing sources, supplying data and adding quotations: up to 40% more visibility in answers.
- schema.org
The vocabulary structured data is validated against, which two of the checks counted here come from.
- RFC 9309
The robots.txt standard, which the three crawler checks come from.
- llms.txt
The index-file convention for agents, still without a settled specification.
Calibration study
Besides the scans people request, we measured a sample of our own to calibrate the methodology thresholds. These figures come from that sample and are published separately, because we chose the websites ourselves and that changes what the numbers mean.
321 domains from Spain, the United Kingdom, the United States and the rest of Europe, spanning brands, services and mid-sized companies, measured on August 4, 2026 with methodology 0.7.0 · provisional thresholds. Figures are calculated over the 226 that could be measured in full.
51 could not be reached: their firewall blocked the analysis. That is 16% of the attempts that got anywhere. Another 41 did not respond or had no content to analyse.
61
Average score across the sample
How are the sample scores distributed?
The sample was put together by hand, sector by sector, so this distribution describes this set of websites and no other. It is not a snapshot of the web: it is a snapshot of the websites we picked in order to calibrate, which is a rather different thing and worth not confusing.
- Ready — no structural blockers
- 9 · 4%
- Citable with gaps — it works, with specific blind spots
- 124 · 55%
- Partially readable — they find you, but cannot cite you reliably
- 80 · 35%
- Not citable — models cannot use this site as a source
- 13 · 6%
What fails most
Share of websites that do not pass each check, over those that could be measured. Beside it, how many websites that share covers.
- A declared stance towards the crawlers that cite you
- 97% · 215/222
- Markdown content negotiation
- 96% · 217/226
- Direct answers after each heading
- 95% · 194/204
- Structured frequently asked questions
- 94% · 212/226
- Entity and business types declared
- 91% · 206/226
- Link discovery headers
- 85% · 192/226
- llms.txt file
- 82% · 182/222
- Paragraphs that stand outside their context
- 66% · 149/226
What does almost the whole sample get right?
These are the checks almost nobody fails. They are a counterweight to the list above: publishing only what fails would paint a darker picture than the real one, and a methodology that shows only the bad news looks too much like a sales argument.
- The crawler is not blocked
- 100% · 226/226
- Time to first byte
- 95% · 215/226
- The crawlers that cite you are not blocked
- 94% · 209/222
- robots.txt reachable and parseable
- 91% · 202/222
- Same content for bots and for people
- 88% · 172/195
- Brand name consistency
- 85% · 160/188
- Identifiable company or author page
- 80% · 181/226
- Clean redirect chain
- 79% · 179/226
The sample was put together by hand, sector by sector, so it is not representative of the web at large: it describes this set of websites and no other. The list of domains is published in the project repository, and no domain appears here with its score.
Have the thresholds moved since the study?
The same sample was measured again on September 22, 2026 with methodology 0.16.0, nine versions after the study above. It is the only way to tell calibrated thresholds from thresholds nobody has looked at, and the first of the two measurements the methodology needs before it stops being a draft.
- 60
- Average of the 206 sites measured without a firewall block — the figure comparable with the study above, which excluded them too
- 53
- Average of all 247 measured, including the 41 behind a firewall
The second figure is lower and the web has not got worse: what changed is who goes into the count. Until version 0.15.0 a site whose firewall blocked the scan was left out of the sample; since then it is measured, and those 41 average 15 because almost nothing about them can be read. Putting that number next to the study without saying this would be exactly the misleading comparison this site holds against others.
On the comparable figure the thresholds have barely moved: 60 now against 61 in the study, across nine versions and seven weeks. That is evidence, not proof — the samples are close but not identical. The measurement that decides is the second one, sixty days from this one and against this same baseline.
Do the tools that sell AI visibility practise it on their own site?
We measured the main website of 12 tools that sell visibility in generative engines, with the same index we apply to everyone else. The average is 73 out of 100, and 8 of the 12 sites do not reach the top band.
Sample chosen by us: the main website of 12 products whose stated purpose is measuring or improving how AI engines cite a site. One page per company, measured on September 21, 2026 with methodology 0.15.0 · provisional thresholds. Every one answered; none blocked the scan.
No names, and that is deliberate rather than an oversight. The index measures a company's website, not its product: a tool can be excellent and still have a marketing site that scores badly, so publishing a name next to a number would imply something this measurement does not support. The scan is free and open to anyone, so anybody can run it on any of them and reach their own conclusion.
73
Average across the sample
How are the 12 sites distributed?
Same bands as any report. The distribution is the finding: the category clusters in the middle, which is where a site works and still leaves specific blind spots.
- Ready — no structural blockers
- 4 · 33%
- Citable with gaps — it works, with specific blind spots
- 6 · 50%
- Partially readable — they find you, but cannot cite you reliably
- 2 · 17%
- Not citable — models cannot use this site as a source
- 0 · 0%
What do they fail most?
Percentage of the sample that fails each check. The order is the measured one, not a selection: the two at the top carry little weight and the ones below them carry a lot, and both facts belong here.
- Link discovery headers
- 83% · 10/12
- Markdown content negotiation
- 75% · 9/12
- Direct answers after each heading
- 50% · 6/12
- Structured frequently asked questions
- 50% · 6/12
- Readable contact details
- 50% · 6/12
- A declared stance towards the crawlers that cite you
- 42% · 5/12
A fixed snapshot, not a live figure. It is not recalculated: repeating the measurement under a later version would be a change of criteria, not a change in the market, so a new round is published as a separate entry. The sample is not disclosed by name for the reason stated above.