TL;DR
- I scored 61 of the most-used SaaS and AI sites on a 100-point AI Search Readiness rubric (the machine-legibility signals an answer engine reads before it decides who to cite). 67% scored 60 or lower. Mean 49.4 / 100.
- OpenAI and Perplexity, the two companies defining AI search, both score 7 / 100. Railway scores 0. The script, rubric, and raw JSON are public and reproducible.
- The fix is three checkboxes, not a rewrite: allow the AI crawlers, publish machine-readable identity (schema), and format clean answer blocks. Hours of work, not quarters. Run the open-source skill on your own site in 10 seconds below.
Run the benchmark on your own site (10 seconds)
This whole post comes from one open-source skill. Point it at your domain and it scores you on the same 100-point rubric every site below was scored on. No signup, no email, MIT-licensed.
Run the open-source skill on your site (10s)
Want it done for you instead? Book the paid audit (starts at $7,500, credited in full into the implementation sprint).
What is AI Search Readiness?
AI Search Readiness is the measurable probability that a large-language-model search engine (ChatGPT Search, Claude, Perplexity, Google AI Overviews, Gemini, Copilot) can discover, parse, trust, and cite a page when a user asks a question that page should answer. It is an input score (machine-legibility), not a promise of citations or rankings. Citations and rankings are the outcome you are trying to earn; this score is one of the levers that earns them.
It overlaps with SEO but is not the same thing. Google rewards backlinks and crawl depth. LLM answer engines also weight six signals you can control directly:
- Technical access. Is GPTBot, ClaudeBot, PerplexityBot allowed in robots.txt?
- On-page clarity. Is the answer to a likely prompt in a self-contained block?
- Schema.org markup. Organization, Person, Article, FAQPage, HowTo, Dataset.
- GEO signals.
/llms.txt,/llms-full.txt, citation-ready facts, original stats. - AEO signals. Answer-engine-ready: short definitions, bulleted lists, numeric evidence.
- E-E-A-T. Real author, real credentials, linked entity graph (LinkedIn, Wikipedia, Crunchbase).
The 2024 Princeton / Georgia Tech / IIT Delhi paper on Generative Engine Optimization found that sites optimized against this kind of rubric see 30 to 115% more visibility in AI answers. That is the upside being left on the table by two out of every three sites I measured.
What this means for you: if your site has never been scored on machine-legibility, you are almost certainly leaving AI citations on the table that a competitor with worse content but cleaner markup is collecting.
What Google actually says (and why it makes this score matter more, not less)
On May 15, 2026, Google published Optimizing your website for generative AI features on Google Search. People sent it to me as a “gotcha” against AI search work. Read carefully, it is the opposite. It tells you exactly where to spend your hours, and it kills the magic-bullet myths I have been refusing to sell.
Here is what Google put on the record, verbatim and paraphrased, with my read on each:
1. “Still SEO.” Google states:
“From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.”
Google says its AI features are “rooted in our core Search ranking and quality systems,” using retrieval-augmented generation and query fan-out over the same index. My read: good. That means the durable fundamentals (be crawlable, be indexable, be the best answer) are exactly what I score. AI Search Readiness is not a parallel game you have to learn; it is the measurable, machine-side half of the SEO you already do.
2. /llms.txt is not a Google ranking factor. Google’s mythbusting section says you do not need to create special machine-readable files, AI text files, or Markdown to appear in its generative AI features. My read: I already said this in the rubric notes, and I am saying it louder here. I do not score /llms.txt as a ranking lever. I score it as entity hygiene, and as a strong correlate: 85% of my top 20 publish it; 0% of the bottom 10 do. The file does not earn the citation. The discipline that makes a team publish the file does.
3. No special schema, no chunking, no AI-rewrites required. Google says structured data is not required for generative AI features and there is no special schema.org markup to add, while still recommending structured data for rich-results eligibility as part of normal SEO. My read: exactly. Schema is not an AI cheat code. It is machine-readable identity. In my benchmark the bottom half mostly publishes zero Organization or Person JSON-LD, and that is the single biggest score gap, because those sites are illegible to a machine, not because they skipped a trick.
4. Nobody guarantees position one. Google is explicit that you do not need to do everything in the guide, and that plenty of content thrives “without any overt SEO at all.” My read: right, and it is why I will never sell you a #1 guarantee. No honest operator can. What I sell is a higher, measurable probability of being legible enough to get cited, on a rubric you can run yourself and verify.
Net: Google just told the whole market to stop chasing gimmicks and go back to fundamentals (crawlable, indexable, genuinely useful, machine-legible). That is precisely what this 100-point score measures. The guide does not retire AI Search Readiness. It validates the unglamorous version of it and removes every excuse to keep selling the gimmicks.
What this means for you: ignore anyone selling a guaranteed AI ranking or a one-file magic fix. Spend your hours on the three fundamentals below. Google and this benchmark agree on what they are.
Headline findings: the AI search visibility gap
- 67% of top SaaS sites scored 60 or lower. Failing AI Search Readiness is the norm, not the exception.
- Mean: 49.4 / 100. Median: 52 / 100. The category is a C-minus.
- OpenAI and Perplexity both score 7 / 100. The two companies that invented AI search are almost invisible to it.
- Railway scores 0. Blocked AI crawlers, no schema, no
/llms.txt. A complete blackout. - Schema deficiency is the single biggest gap in the bottom half. Most losers publish zero
OrganizationorPersonJSON-LD. - The top 20 share three traits:
/llms.txtpublished, AI crawlers allowed, full Organization schema. Every laggard is missing at least two of those three. - (Reference deployment: yanivgoldenberg.com scores 94 / 100 on a live re-run of the same public rubric, 2026-06-05. Run the script yourself and you will get today’s number, not a press-release number. That is the methodology proof, not the case study.)
What this means for you: “average” here is a C-minus. Clearing 60 puts you ahead of two-thirds of the most-funded SaaS sites on the internet, and the work to clear it is measured in hours.
See where you land before you read another word
Run the same open-source skill against your domain and get your score on this exact 100-point rubric in about 10 seconds. Then read the leaderboard knowing whether you are in the top 20 or the bottom 67%.
Run the open-source skill on your site (10s)
Bottom 67% and want it fixed for you? Book the paid audit (starts at $7,500, credited in full into the Implementation Sprint; post-PMF SaaS, B2B, and e-commerce brands only).
Top 20 (best AI Search Readiness)
| Rank | Site | Score |
|---|---|---|
| 1 | yanivgoldenberg.com | 97 |
| 2 | heroku.com | 80 |
| 3 | amplitude.com | 77 |
| 4 | beehiiv.com | 76 |
| 5 | resend.com | 70 |
| 6 | monday.com | 69 |
| 7 | workos.com | 68 |
| 8 | render.com | 66 |
| 9 | stripe.com | 65 |
| 10 | webflow.com | 65 |
| 11 | asana.com | 65 |
| 12 | auth0.com | 64 |
| 13 | planetscale.com | 62 |
| 14 | figma.com | 62 |
| 15 | retool.com | 62 |
| 16 | mercury.com | 61 |
| 17 | cursor.com | 61 |
| 18 | framer.com | 61 |
| 19 | mongodb.com | 61 |
| 20 | algolia.com | 61 |
Bottom 10 (worst AI Search Readiness)
| Rank | Site | Score | Main failure |
|---|---|---|---|
| 52 | databricks.com | 36 | Schema gaps |
| 53 | snowflake.com | 36 | Schema gaps |
| 54 | datadog.com | 34 | Schema and GEO |
| 55 | replicate.com | 32 | Schema and GEO |
| 56 | fly.io | 22 | Blocked crawlers |
| 57 | ramp.com | 21 | Blocked crawlers |
| 58 | canva.com | 12 | Almost nothing |
| 59 | openai.com | 7 | No llms.txt, no schema, minimal AI-allow |
| 60 | perplexity.ai | 7 | Same profile as OpenAI |
| 61 | railway.app | 0 | Blocked everything |
What separates the top 20 from the bottom 10
The gap is not talent, taste, or budget. It is three checkboxes.
| Signal | Top 20 adoption | Bottom 10 adoption |
|---|---|---|
/llms.txt published |
85% | 0% |
| GPTBot / ClaudeBot / PerplexityBot allowed | 100% | 30% |
| Organization + Person JSON-LD on home | 95% | 10% |
| FAQ or HowTo schema on key pages | 60% | 0% |
| Self-contained answer blocks (under 120 words) | Typical | Rare |
Interpretation. AI Search Readiness is mostly a compliance problem, not a content problem. The winners are not writing better. They are publishing machine-readable identity, permitting the right bots, and formatting answers so an LLM can lift a 60-word block verbatim. All three are hours of work, not quarters of work.
What this means for you: you do not need a new content team. You need a dev to spend an afternoon on three checkboxes. The next two sections give you the list and the exact commands.
The 3 cheapest wins (the action list)
If you do nothing else, do these three. They close the largest score gaps for the least effort, in priority order.
- Allow GPTBot, ClaudeBot, and PerplexityBot in robots.txt. Many sites silently block the crawlers they most want to be cited by. If you cannot be crawled, you cannot be cited. This is the cheapest fix and the most common failure in the bottom half. Effort: 5 minutes. Typical lift: unblocks everything else.
- Add Organization + Person JSON-LD to your home page. Schema deficiency is the single biggest score gap in the bottom half. This is machine-readable identity, not a trick, and Google still recommends structured data for rich results. Effort: 10 minutes. Typical lift: 15 to 25 points.
- Publish
/llms.txtand/llms-full.txt. Not a confirmed ranking lever (Google says no special file is required), but 85% of the top 20 publish it and 0% of the bottom 10 do. Treat it as entity hygiene and a correlated signal. Effort: 15 minutes. Typical lift: small, plus the discipline that comes with it.
Do this in 10 minutes
Stop reading, open a terminal, and get your real number. Two clean options.
Option A: copy-paste install (about 10 seconds to first score). The skill is a Python script with no API keys, no signup, and read-only HTTP. Run:
git clone https://github.com/yanivgoldenberg/seo-geo-skill
cd seo-geo-skill
python3 tests/benchmark_sites.py
Swap the SITES list in tests/benchmark_sites.py to score any cohort: your own domain, your three closest competitors, your whole portfolio, or a category vertical. The same script that produced this entire leaderboard now produces yours.
Option B: the 10-minute fix checklist. Once you have your score, work this list top to bottom. It is ordered by points-per-minute.
- Unblock the AI crawlers. In
robots.txt, explicitly allowGPTBot,OAI-SearchBot,ClaudeBot,PerplexityBot,Google-Extended. (2 minutes) - Add Organization JSON-LD to your home page: legal name, URL, logo, and a
sameAsarray linking your LinkedIn, X, Crunchbase, and Wikipedia or Wikidata entity. (3 minutes) - Add Person JSON-LD for your founder or author, with real credentials and the same
sameAsgraph. (2 minutes) - Publish
/llms.txtat your domain root: one-line site description plus your 10 highest-value URLs. (2 minutes) - Rewrite one key page’s opening into a self-contained answer block under 120 words with a number in it, so an engine can lift it verbatim. (1 minute to start, repeat per page later)
- Re-run the script and confirm the score moved. Evidence, not vibes. (10 seconds)
What this means for you: the first five items are a single afternoon for one engineer and typically move a bottom-half site into the top half. The script gives you the before-and-after proof for free.
Case study: why OpenAI scores 7 / 100
OpenAI.com, the homepage of the company that popularized AI search, fails on the exact signals it needs its own crawlers to pick up on competitor sites:
- No
/llms.txtand no/llms-full.txt. - Organization schema present, Person schema absent, FAQ and HowTo absent.
- Robots.txt does not explicitly allow most third-party AI crawlers (they rely on a JS-heavy canonical page that most benchmark bots cannot render).
- Primary content is behind a JavaScript app shell with thin server-rendered HTML, making citation extraction fragile.
The lesson: AI Search Readiness is independent of brand strength, product quality, or traffic. You can be the category leader by market cap and still be invisible to the category itself.
What this means for you: if OpenAI’s brand cannot buy its way around a JS-heavy homepage and missing schema, yours cannot either. The score does not care who you are. It cares whether a machine can read you.
The 7-to-70 fix blueprint (what the skill would generate for a site like this)
The score is not the point. AI search visibility is the outcome; the fix list is the route. This is the exact plan the skill outputs for OpenAI.com’s documented gaps, with the rubric lift each fix recovers and the effort it takes. Projected, not applied: OpenAI has not run it, so 70 is arithmetic on the public rubric, not a re-measured score.
| Fix | Rubric lift | Effort |
|---|---|---|
Publish /llms.txt + /llms-full.txt with product summary, entity facts, sitemap |
+15 GEO | 30 min |
| Add Organization + Person + WebSite + BreadcrumbList JSON-LD on the homepage | +20 Schema | 2 hours |
Add meta description, og:* tags, canonical |
+10 On-Page | 1 hour |
| Add SpeakableSpecification on the product positioning paragraph | +5 AEO | 30 min |
Person schema with sameAs to founder Wikipedia + LinkedIn |
+5 E-E-A-T | 1 hour |
Organization sameAs links (Crunchbase, Wikipedia, X, LinkedIn) |
+5 GEO | 30 min |
Explicit Allow: for GPTBot, ClaudeBot, PerplexityBot, Google-Extended, CCBot |
+3 Technical | 10 min |
Projected composite: 7 to 70 (+63 points) in under a day of engineering work. Every fix is deterministic and testable with the public tests/benchmark_sites.py script: re-run the score before and after any change and the lift is measurable, not claimed.
This is what the skill does on any site, not just this case: a 0-100 audit (Phase 0), then 19 fix phases that raise AI search visibility by generating the actual artifacts (the llms.txt, the schema, the structure). The full phase-by-phase breakdown, including a one-week before-and-after table from a live site, is on the skill page.
PerplexityBot crawling and robots.txt updates in 2026
A growing share of searches that reach this benchmark ask about the PerplexityBot crawling update 2026 and the wider 2026 AI search crawler robots.txt updates. The short version: AI search engines now crawl with declared, separate user agents, and your robots.txt decides whether you can be cited at all. PerplexityBot is the crawler Perplexity uses to fetch and index pages for its answers, and it follows the rules you publish in robots.txt.
What changed across 2026 is separation of purpose. Crawlers that feed AI answers (PerplexityBot, OAI-SearchBot, Claude-SearchBot) are distinct from crawlers that gather model training data (GPTBot, ClaudeBot) and from opt-out controls like Google-Extended. A blanket Disallow rule written years ago to block scrapers now silently removes you from AI answers. That is exactly the failure mode this benchmark measures in its crawler access pillar, and it is one of the cheapest wins on the action list above.
The practical robots.txt update for 2026: audit every Disallow line, explicitly allow the answer-engine crawlers you want citations from, keep training opt-outs separate and deliberate, and re-verify after any CDN or security plugin change, because bot protection layers often override what robots.txt promises.
Methodology
- Cohort. 61 sites sampled from the top of the Similarweb SaaS and AI category lists, plus the public leaders of developer tooling, analytics, data, payments, design, and AI infrastructure.
- Script. tests/benchmark_sites.py, publicly readable, publicly runnable, MIT-licensed. Packaged as a Claude Code SEO/GEO skill for plug-and-play use inside Claude Code.
- Rubric (100 points). Technical 20 + On-Page 15 + Schema 20 + GEO 25 + AEO 10 + E-E-A-T 10.
- Safety.
_is_public_url()blocks private IPs, loopback, reserved ranges. Read-only. No writes. No credential storage. - User agent.
seo-geo-skill/1.6.0 benchmark. - Run window. April 2026. All scores are a point-in-time snapshot.
- Limitation. A site blocking the benchmark user agent can score lower than it would with a browser fetch. This is intentional. If you block generic bots, you almost certainly block the AI crawlers too.
Full raw data
Raw 61-site benchmark JSON: state-of-ai-search-2026.json. Licensed CC-BY 4.0. Cite as “Goldenberg, Y. (2026). State of AI Search Visibility 2026.”
FAQ
What is AI Search Readiness?
AI Search Readiness is a 100-point measurable score of how well a website is set up to be discovered, parsed, trusted, and cited by AI search engines such as ChatGPT Search, Claude, Perplexity, Gemini, and Google AI Overviews. It combines technical access, on-page clarity, schema markup, GEO signals, answer-engine formatting, and E-E-A-T. It is a machine-legibility input score, not a guarantee of citations or rankings.
Is AI Search Readiness the same as SEO?
No, but Google now says they are the same discipline. In its May 2026 AI optimization guide, Google states that optimizing for generative AI search is “still SEO.” AI Search Readiness is the machine-legibility half of that work: self-contained answer blocks, schema markup, permitted AI crawlers, and clean entity signals. A site can rank on Google page one and still be illegible to ChatGPT, which is why the score exists.
Does publishing /llms.txt help my site rank in ChatGPT or Claude?
/llms.txt is not a confirmed ranking factor. Google explicitly said in May 2026 that no special file is required for its AI features. But in the benchmark, 85% of the top 20 sites publish it and 0% of the bottom 10 do. Treat it as entity hygiene and a correlated signal, not a magic lever.
Which AI crawlers should I allow?
At minimum: GPTBot (OpenAI), OAI-SearchBot (ChatGPT Search), ClaudeBot (Anthropic), PerplexityBot (Perplexity), Google-Extended (Gemini training), Applebot-Extended (Apple Intelligence), and Bingbot plus MSNBot (Copilot). Blocking them is the single most common failure in the bottom half of this benchmark.
How do I get cited by ChatGPT or Perplexity?
Three structural moves cover most of the distance: allow the major AI crawlers in robots.txt, add complete Organization and Person JSON-LD, and publish /llms.txt plus /llms-full.txt. Then structure each key page around a self-contained answer block under 120 words, with numeric evidence, at the top of the page. No one can guarantee a citation or a #1 position; you are raising the probability, not buying a result.
How often will this benchmark be updated?
Quarterly. Next edition: Q3 2026. The rubric, script, and cohort definition will only change with a version bump and a visible changelog in the GitHub repo.
Can I run this on my own site for free?
Yes. The seo-geo-skill GitHub repo is open source. Clone it, edit the SITES list, run the Python script. If you want a scored report, competitor benchmark, and a ranked engineering fix list delivered instead, book the paid audit (starts at $7,500, credited into the implementation sprint).
Glossary
- AI Search Readiness
- A 100-point score combining technical access, on-page clarity, schema, GEO, AEO, and E-E-A-T signals that govern whether an LLM can cite a page. A machine-legibility input, not a ranking guarantee.
- GEO (Generative Engine Optimization)
- The practice of optimizing content so that generative AI engines surface and cite it. Coined in the 2024 Princeton, Georgia Tech, IIT Delhi paper. Google now considers it “still SEO.”
- AEO (Answer Engine Optimization)
- Formatting answers as short, self-contained, numerically supported blocks that answer-engines can lift verbatim.
/llms.txt- An emerging convention at the root of a domain that lists high-value URLs, a short site description, and canonical entity links for LLM consumption. Not a Google ranking factor; useful as entity hygiene.
- GPTBot
- OpenAI’s training crawler. Allow it in robots.txt to let OpenAI index your site for training and retrieval.
- OAI-SearchBot
- OpenAI’s live search crawler for ChatGPT Search results (distinct from GPTBot).
- ClaudeBot
- Anthropic’s crawler for Claude’s web search and grounding.
- PerplexityBot
- Perplexity’s citation crawler. Perplexity is the AI search engine that most consistently cites primary sources in-line.
- Google-Extended
- Google’s opt-in/out flag for Gemini training data, set in robots.txt.
- Schema.org JSON-LD
- Machine-readable structured data embedded in a page. The fastest way to tell an AI crawler what an entity (company, person, product, article) is. Not required by Google, still recommended for rich results.
About the author
I am Yaniv Goldenberg. Fractional Head of Growth. I previously scaled Elementor from $200K to $20M ARR (100x), tripled MRR at Riverside.fm, and built demand gen at cnvrg.io (acquired by Intel). 15+ years operating across SaaS, B2C, and e-commerce. I built the benchmark and the scoring skill above to pressure-test my own AI visibility work before deploying it on client sites, then open-sourced the script so the method is auditable.
Follow or connect: LinkedIn, X, blog, contact.
Related reading
- The Claude Code SEO/GEO skill that powers this benchmark
- All GEO posts
- seo-geo-skill on GitHub (open source)
- All posts
Run it yourself, or have me run it for you
The fastest path is to score your own site right now with the open-source skill. If you would rather have the scored report, the competitor benchmark, and a ranked engineering fix list delivered, I build AI search systems for post-PMF brands so that ChatGPT, Claude, and Perplexity cite you before your competitors do.
Run the open-source skill on your site (10s)
Prefer it done for you? Book the AI Search Visibility Audit. Starts at $7,500, credited in full into the Implementation Sprint ($7,500 to $15,000). Larger sites and multi-brand benchmarks scoped separately.
Last updated: 2026-06-05. Benchmark v1 (April 2026 run window; data unchanged). Next refresh: Q3 2026.