momentence

Original research, fielded 2026-08-09

The AI Discovery Technical Hygiene Benchmark 2026

We crawled 160 live software company websites and scored the 10 technical hygiene signals AI crawlers and parsers depend on. Median score: 7 out of 10. Only 8 percent publish FAQ schema. This study measures access and legibility on a company's own domain, not whether assistants cite it. Methodology and raw data below, free to cite.

Read the methodology

160

Software sites analyzed

175 attempted, 15 excluded

7/10

Median technical hygiene score

Mean 6.33

52%

Serve a real llms.txt

Verified as plain text, not a soft 404

81%

Have no AI crawler policy

Zero of 11 AI agents named in robots.txt

The short answer

Software websites are well optimized for humans and poorly optimized for the machines that now summarize them. Across 160 live sites, the median technical hygiene score was 7 out of 10, 34 percent scored 8 or higher, and 18 percent scored 4 or lower. The gap is concentrated in structured data and explicit crawler policy, which are the two cheapest things on the list to fix.

  • Classic on-page hygiene is largely solved: 95% have a meta description and 90% a canonical tag
  • Machine-facing hygiene is not: 38% publish no JSON-LD and only 8% publish FAQPage schema
  • 81% of sites do not address a single AI crawler in robots.txt, and only 5% block any
  • Hygiene tracks category maturity, from a mean of 7.1 in B2B horizontal SaaS to 4.5 in healthtech

The data

What we found

Every figure is the share of the 160 analyzed sites where the signal was present in the raw HTTP response on 2026-08-09.

Signal presence across 160 software homepages (percent of sites)
Meta description present
95%
Open Graph title present
93%
Valid robots.txt
92%
Canonical tag
90%
Sitemap declared in robots.txt
86%
sitemap.xml at root
79%
Exactly one H1
78%
Any JSON-LD on homepage
62%
Organization entity schema
56%
llms.txt at root (adoption only, no documented effect)
52%
WebSite schema
32%
Product or SoftwareApplication schema
17%
BreadcrumbList schema
14%
FAQPage schema
8%
Distribution of technical hygiene scores
2
1
3
2
11
3
12
4
27
5
17
6
34
7
39
8
14
9
1
10

Technical hygiene score out of 10 on the horizontal axis, number of sites on the vertical axis. n = 160.

Common on-page defects (percent of sites)
Multiple H1 elements
14%
No H1 at all
8%
Title over 60 characters
27%
Meta description over 160 characters
26%
AI crawler policy in robots.txt (percent of sites, by agent)
Explicitly allowedBlockedNot addressed at all
GPTBot
ClaudeBot
Google-Extended
PerplexityBot
ChatGPT-User
OAI-SearchBot
CCBot
Applebot-Extended
anthropic-ai
meta-externalagent
Bytespider

Technical hygiene by category

Technical hygiene by software category
CategorySitesMean scorellms.txtJSON-LDFAQPage
B2B horizontal SaaS367.167%83%11%
Developer tools and infrastructure396.767%72%3%
Fintech and payments276.659%63%11%
Martech and analytics196.653%58%5%
Consumer software265.127%38%15%
Healthtech134.58%31%0%

Median homepage HTML weight was 446.1 KB, mean 556.8 KB, 90th percentile 1116.6 KB. Median title length was 53 characters and median meta description 148 characters.

Analysis

Six findings that matter

01

Structured data is the biggest single gap

38 percent of software homepages carry no JSON-LD at all, and only 8 percent publish FAQPage schema. Structured data removes ambiguity for any machine parsing the page and makes a site eligible for certain search features. Google states that no special AI schema is required, so we do not present it as a citation lever. We present it as the cheapest way to stop asking machines to infer facts you could simply declare.

02

llms.txt adoption is widespread and unproven

52 percent of the full sample serves a real llms.txt, with a stark split: 67 percent in B2B horizontal SaaS and developer tools versus 27 percent in consumer software and 8 percent in healthtech. Adoption is not evidence of effect. Google has said publicly that it does not use llms.txt, and no major provider documents it as a ranking or citation input, so the honest reading is that teams in AI-influenced categories are hedging cheaply. Treat it as an experiment to measure, not a foundation.

03

Almost nobody has an explicit AI crawler policy

81 percent of sites do not mention a single one of the 11 AI crawlers we checked in robots.txt, and only 5 percent block any of them. Non-blocking is the right default for a company that wants to be cited, but silence is not a decision. Teams inherit whatever each vendor's default behavior is, then discover the consequences after the fact.

04

Human-facing hygiene is solved, machine-facing hygiene is not

The human-facing basics are nearly universal: 95 percent have a meta description, 92 percent a valid robots.txt, 90 percent a canonical. The machine-facing layer collapses right after that: 56 percent Organization schema, 17 percent product schema, 14 percent breadcrumbs. Two decades of SEO discipline produced pages that read well and describe themselves poorly.

05

Homepages are heavy enough to hurt retrieval

Median homepage HTML is 446 KB, mean 557 KB, and the 90th percentile clears 1.1 MB. Heavy client-rendered shells raise the odds that a crawler with a modest budget or no JavaScript execution sees a fraction of what a browser does. Content that never reaches the parser cannot be cited.

06

Category maturity, not company size, predicts hygiene

Mean hygiene score runs 7.1 in B2B horizontal SaaS down to 4.5 in healthtech, and several of the lowest scores in the sample belong to very large, very well-funded companies. Hygiene tracks whether a category's buyers already research through AI assistants, not headcount or revenue.

Methodology

How the study was run

Published in full so the results can be audited, reproduced, or argued with.

Collection

  • Fielded on 2026-08-09 from a single crawler identifying itself as MomentenceBenchmarkBot.
  • 175 real, currently operating B2B software and SaaS domains were attempted across 6 categories. 160 returned a usable homepage response and 15 were excluded for timeouts, blocks, or non-200 responses.
  • For each domain we fetched the homepage, /robots.txt, /llms.txt, and /sitemap.xml.
  • We parsed the raw HTML response and did not execute JavaScript, because retrieval crawlers frequently do not either.
  • robots.txt was parsed into user-agent groups, and each of 11 AI agents was classified as blocked by a blanket disallow, explicitly named and allowed, or not addressed.
  • llms.txt results were verified a second time to reject soft 404s, HTML fallbacks, and empty stub files. Three apparent hits failed verification and were reclassified as absent.

Scoring

One point per signal present, out of ten, unweighted:

  1. 01Any valid JSON-LD block on the homepage
  2. 02Organization or equivalent entity schema
  3. 03FAQPage schema
  4. 04Product, Service, or SoftwareApplication schema
  5. 05llms.txt served as plain text at the root (included to measure adoption; Google has stated it does not use the file, so read this component as a habit signal, not a requirement)
  6. 06sitemap.xml resolving to a valid urlset or sitemapindex
  7. 07Meta description present
  8. 08Exactly one H1 element
  9. 09Canonical tag present
  10. 10No blanket disallow against any of the 11 AI crawlers checked

Limitations, stated plainly

This is a purposive judgment sample of recognizable software companies, not a random sample of the software industry, so the figures describe the visible end of the market and are not statistically projectable to all software companies. Only the homepage was scored, so a site with excellent schema on its product and comparison pages can score low here. The crawl is a single point in time and reflects the response served to one non-browser user agent. Most importantly, technical hygiene is a precondition for being used as a source, not proof of citation: nothing in this dataset measures whether an assistant cited, mentioned, or accurately described anyone. Measuring that requires querying the assistants directly against a fixed prompt set. llms.txt is included as an adoption measure only; Google has stated it does not use the file, so its presence in the score should be read as a habit signal rather than a requirement.

Data, licence, and citation

The full per-domain dataset, including every signal and the computed score, is published as CSV. It is free to reuse, quote, and republish under CC BY 4.0 with attribution to Momentence and a link to this page.

Momentence. AI Discovery Technical Hygiene Benchmark 2026. Fielded 2026-08-09. https://www.momentence.com/reports/ai-discovery-technical-hygiene-benchmark

What to do about it

Five ways teams get this wrong

If you want the fix list rather than the failure list, the definitive SEO, AEO, and GEO guide has the checklists, and answer engine optimization is where this work sits inside a Momentence program. Our methodology page shows the month-by-month sequence we run it in.

FAQ

Questions about this benchmark

For each domain we fetched the raw homepage HTML response, robots.txt, llms.txt, and sitemap.xml, then parsed ten binary technical hygiene signals from the response bodies. We did not execute JavaScript, because the signals we care about are the ones a retrieval crawler can read without rendering. Every measurement is a fact about a real HTTP response captured on 2026-08-09.
We assembled a judgment sample of 175 real, currently operating B2B and B2C software companies across six categories, then analyzed the 160 that returned a usable homepage response. Fifteen domains were excluded for timeouts, blocks, or non-200 responses. This is a purposive sample of recognizable software companies, not a random sample of the entire industry, so the results describe the visible end of the market rather than a statistically projectable universe. We say so plainly because the alternative is a false precision claim.
One point per signal present, out of ten, unweighted. Unweighted is a deliberate choice: any weighting scheme would encode our opinion about which signal matters most, and we would rather publish the raw components so you can weight them yourself. The per-domain components are all in the CSV.
Yes. The findings, charts, and dataset are free to cite, quote, and republish with attribution to Momentence and a link to this page. If you rerun the crawl and get different numbers, publish them. Benchmarks are only useful when they are contestable.
Yes, on the same methodology and the same sample frame, so the series stays comparable. Each edition keeps its own dated dataset rather than overwriting the last one.
No, and we will not claim it does. This benchmark measures technical hygiene: whether crawlers can reach your pages and whether machines can parse them. That is a precondition for being used as a source, not a cause of citation. Whether an assistant cites or accurately describes you also depends on content quality, topical authority, and third-party sources, none of which a crawl of your own domain can observe. Google states that its search fundamentals still apply to AI experiences, which is roughly what this study measures.

Score your own site against the benchmark

We will run the same ten-signal audit on your domain, send you the raw output alongside the benchmark medians, and tell you which two fixes actually move you. No obligation.