GEO Optimizer runs 16 CLI commands, 47 content checks, and 27 AI crawler tests per audit. After 1,800+ audits and a calibration study that came back null, I have an honest ranking of which checks matter and which are table stakes. Here is that ranking.
Tier 1: These checks predict something real
These are the checks where a failure means you are almost certainly invisible to at least one AI engine. Fix these first.
robots.txt blocking AI crawlers. The audit checks 27 AI crawlers (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Claude-Web, and 22 more). If any are blocked, that engine cannot read your page. 9.5% of audited sites block at least one AI bot in robots.txt. On the subset where the audit also tests HTTP-level blocking (about 630 sites), the combined rate rises to roughly 20%, with CDN/WAF-level blocking at least as common as explicit Disallow rules. Many of these blocks come from legacy security plugin rules (Wordfence, Sucuri) that block unknown user agents by default. This is the single most common critical issue. If GPTBot is blocked, ChatGPT's 14.4% web search rate cannot reach you. If all 27 are blocked, you are invisible to every AI engine's live retrieval.
JavaScript-only content. The audit checks whether the main content is in the raw HTML or rendered client-side. If the page is an SPA that renders content via JavaScript, AI crawlers see an empty shell. The HN crawler study of 1,500 sites found 40% of marketing sites rely so heavily on client-side rendering that they appear as empty shells to non-hydrating AI agents. This is not a score issue. It is a visibility issue. If the crawler sees nothing, every other check is irrelevant.
Canonical URL missing or mismatched. The audit checks whether the page has a canonical URL and whether it matches the requested URL. The WashU study (55,393 queries) found that canonical correctness correlates with AI Overview citation. A missing or mismatched canonical means the AI engine may not attribute the content to the right URL, which means it may not cite the right page.
Tier 2: These checks matter but are not critical
These checks contribute to the score and align with what the data says about citation. A failure here means you are missing a signal, not that you are invisible.
Published and updated dates. The audit checks for visible dates on the page. The Babel42 study found that pages ignored by AI Overviews were thinner and less likely to have dates. The WashU study found dates correlate with citation. Missing dates cost you up to 4 points in the Google AI readiness category.
Named author byline. The audit checks for a named author. The WashU study found that credible sources with clear authorship are cited more. A missing byline costs you Google AI readiness points.
Schema markup (Organization, WebSite). 76% of audited sites ship some JSON-LD. The sites that do not are missing a signal that helps AI engines understand what the page is about. Organization and WebSite schema alone cover most of the value. Adding a sixth or seventh schema type has diminishing returns.
Meta description and title tag. Basic SEO. The audit checks for presence and quality. Missing these is table stakes that should have been fixed years ago, but 15-20% of audited sites still miss them.
Tier 3: These checks are honest about their limits
These checks are in the tool because they measure something real, but the calibration study showed they do not predict citation. They measure citability, not citation.
llms.txt presence and validity. Worth 6 points in rubric v2 (down from 18 in v1). 58% of audited sites have one. Google says it does nothing for AI Overviews. Some AI crawlers read it. The calibration study showed no relationship between llms.txt points and Perplexity citation. It is the easiest fix and the lowest-impact one. I keep it in the tool because it is a real signal for some crawlers, but I weight it honestly at 6 points.
Citation signals (quotable passages, statistics, quotes). The audit checks for content patterns that are easy for AI to extract: self-contained answer blocks, statistics with sources, quotable sentences. These align with the Babel42 finding that 67% of AI Overview quotes come from the first 30% of the page. But the calibration study showed no significant relationship between this category and citation. The content structure helps extractability. It does not guarantee citation.
Brand entity clarity. The audit checks whether the page has consistent entity references (same name, same description, same links) across schema, metadata, and content. This helps AI engines disambiguate the brand. The calibration study showed no significant relationship (p = 0.2227, Holm-corrected p = 1.0000). It is still a useful check for brand consistency, but it does not predict citation.
Tier 4: These checks are structural, not predictive
These checks exist because they measure something about the page's technical health. They do not predict citation, and they never claimed to.
AI discovery signals (RSS, sitemap references). Worth 3 points in rubric v2 (down from 6). These check whether the site has discovery mechanisms for AI crawlers. Minor signal. I keep it because it is a real technical check, but it is the lowest-weighted category.
Google AI readiness (8 sub-signals). Worth 20 points. This is the newest category, added in v4.19.0. It checks canonical, dates, byline, links, viewport, sitemap, snippets, and index. The calibration study showed the closest-to-significant result here (p = 0.1220 before Holm correction), trending in the wrong direction (OR = 0.8345). This category is based on the WashU study's findings about what correlates with AI Overview citation. The calibration study tested it against Perplexity, not Google. The relationship may be different on Google AI Overviews. I need to run the study again on Google before I can say whether this category predicts anything.
What I would remove
Nothing. Every check measures something real about the page. The question is how to weight them, not whether to keep them. The calibration study told me the weights do not predict citation. The next step is to run the study on Google AI Overviews, where the source selection mechanism is different from Perplexity.
If the Google study also comes back null, I will need to reconsider whether a 0-to-100 score is the right output. A checklist of "fix this, this, and this" may be more honest than a single number that implies predictive power it does not have.
Try it
If you want to see which tier your site falls into, run a free audit on geoready.dev. The score tells you where you are. The fix list tells you what to do. The tier ranking above tells you which fixes to prioritize.
The CLI is on GitHub, MIT licensed, 1,137 stars, 2,000+ tests. The full scoring rubric is in SCORING_RUBRIC.md. If you disagree with a tier ranking, open an issue. The data is open. The calibration study is open. The disagreement should be too.
Correction (Nov 12): The original version of this article stated 30% of audited sites block at least one AI bot. After rechecking the benchmark data, the robots.txt blocking rate is 9.5% across 2,800 domains. On the subset where HTTP-level blocking is also tested, the combined rate is roughly 20%. The 30% figure was not supported by the data and has been corrected.
Top comments (16)
A large checklist becomes maintainable when each check has a decision attached: block release, create a ticket, or record a baseline. Iād also sample false positives and measure time-to-diagnosis, since a technically accurate test can still be a poor guardrail if nobody trusts or acts on it. The useful suite is the one that changes decisions.
You're right on the principle: each check needs a decision attached, or it's noise. The tier ranking in the article is that mapping. Tier 1 means fix now or stay invisible. Tier 3 means record a baseline, it doesn't predict anything. Tier 4 is structural, no citation prediction claimed.
On false positives and time-to-diagnosis: fair point the article doesn't cover. The calibration study tested the score at the suite level (came back null), but I haven't measured false positive rates for individual checks yet. For example, robots.txt detection has a false positive when a site uses a CDN that intercepts the crawl before robots.txt is served. Worth tracking.
One context difference: the tool is an audit, not a CI gate. It doesn't block deploys. But the principle holds. A check that doesn't change a decision is dead weight. The calibration study was exactly that. It showed the score doesn't predict citation, so the 0-100 number doesn't change decisions as much as the fix list does.
That audit framing makes the split even clearer: the dashboard should optimize for an ordered next action, while the calibration layer tells you how much to trust each check. Logging false positives by delivery pathāorigin, CDN, or WAFāwould also turn those edge cases into routing rules rather than caveats buried in the report.
The tier ranking is already that ordered action list. Tier 1 fixes first, then Tier 2, down the line. The trust layer on top is what's missing.
Logging false positives by delivery path is a concrete improvement. Right now a robots.txt check treats a CDN interception and a real blocking rule the same way. Splitting them by path would let the tool say "probably CDN interception" instead of flagging both as failures. Not built yet, but it's where the check layer needs to go.
That distinction would make the trust layer actionable. Iād retain the observed evidence with each findingārequest path, headers or cache signal, and where the response divergedāthen expose a confidence label and a short remediation path. A CDN-shaped anomaly can then be triaged without pretending it proves a crawler block.
This is a solid data model for the finding layer. Retaining the request path, headers, and where the response diverged gives the confidence label something to stand on. A CDN interception and a real Disallow rule produce different evidence, and right now the tool treats both the same way.
One practical constraint: the tool is a stateless CLI. Capturing request/response metadata per finding means the check layer needs to collect that evidence during the audit, then carry it through to the output. That's a structural change to how checks work today. Feasible, but it belongs in the next rubric revision rather than a patch.
The confidence label is the easy part once the evidence is there. The hard part is collecting that evidence reliably across different CDN and WAF setups without introducing false negatives of its own.
Yesāthe split between observable facts and interpretive guidance is what lets a trust layer support action instead of just scoring pages. Iād expose the evidence timestamp and collection method beside every assertion as well, so a stale crawl does not carry the same weight as a recently verified source. Freshness is part of the claim.
Freshness is a valid point. A stale crawl shouldn't carry the same weight as a recent one.
There's a practical wrinkle with a stateless CLI, though. In a single run, every finding shares the same timestamp because the tool does one pass. Per-finding timestamps would be identical across all results from that run. Where freshness gets useful is continuous monitoring, comparing successive runs over time. A one-shot audit doesn't have that problem because you always know when it ran.
The collection method per finding is more interesting than the timestamp. Whether it was a GET vs HEAD, which UA string was used, whether there was a redirect: those details change how you interpret the result. They're also easier to capture during the run than a timestamp that's the same for everything.
So: timestamp at the run level, collection method at the finding level. Freshness is a monitoring concern, not a one-shot audit concern.
In one scan this week, GPTBot and ClaudeBot got a bot-challenge page from a large site. Of the 30% of sites you found blocking at least one AI bot, how many used robots.txt and how many a WAF or CDN rule?
The tool checks the two mechanisms separately, but the full dataset only stores the robots.txt counts.
The audit reads robots.txt (Disallow rules per bot) and, in a separate step, requests the home page with the GPTBot, OAI-SearchBot, PerplexityBot and Claude-SearchBot user agents, comparing status code and body size with a normal browser request. Only the robots.txt counts are kept for the full 2,800-domain benchmark.
The HTTP check was stored for about 630 sites, mostly registered accounts and one-off audits. On those, 7% block at least one AI bot in robots.txt, and 9% return 403/429/451/503 to an AI user agent while the browser gets a 200. Counting challenge-page keywords and truncated bodies too, 16% look blocked at the HTTP level. Taking the strict definition: 31 sites block only through robots.txt, 45 only at the HTTP level, and 13 do both.
So on this subset, CDN/WAF-level blocking is at least as common as an explicit Disallow. Two caveats: the HTTP check has false positives (it doesn't follow redirects, and a page containing "blocked" or "captcha" can trip it), and it only sees status and body, so it can't distinguish Cloudflare rules from a Wordfence or Sucuri rule. About half of the HTTP flags are on Cloudflare-fronted sites.
I've also corrected the blocking rate in the article above. The original 30% figure was not supported by the benchmark data. The actual rate is 9.5% via robots.txt across 2,800 domains, rising to roughly 20% when HTTP-level blocking is included.
We run a check for that gap. It asks for the page as an AI agent and as a browser, then compares what comes back.
On the Cloudflare half, did the block look like a default for new zones, or a switch the owner turned on?
We can't tell from the data the tool collects today. It sees the outcome (status code, challenge page), not the Cloudflare configuration behind it.
What we know: Cloudflare Bot Fight Mode is off by default on new zones. If it's on, someone turned it on. But Cloudflare also has WAF managed rules and rate limiting that can intercept bots without the site owner explicitly configuring an AI crawler block.
One indirect signal: if the block hits only GPTBot or ClaudeBot specifically, it's more likely a configured rule. If it hits any unrecognized user agent, it's more likely Bot Fight Mode. The tool doesn't split those two cases in its output today, but it's a distinction that could be added by comparing which user agents get blocked on each site.
Our check asks for the page as GPTBot, as ClaudeBot and as a browser. A fourth request under a made-up agent name would sit next to those and show which case it is.
Would you count a block on the made-up name as Bot Fight Mode, or keep it as a third bucket?
Good idea. The fourth request is the cleanest way to split the two cases.
I wouldn't label it "Bot Fight Mode" specifically. From the outside we can't distinguish BFM from Super Bot Fight Mode, from a WAF custom rule blocking non-browser UAs, or from origin-level security plugins (Wordfence, Sucuri) that block unrecognized user agents. All of them produce the same observable result: the made-up UA gets blocked. The test tells us whether the block is generic or specific to AI crawlers, not which mechanism causes it.
Three buckets:
The third one is the case I care about most. It means the site has bot protection active but has chosen to let verified AI crawlers through. For the tool it's a pass on AI visibility, not a block, so it deserves its own bucket.
Agreed on dropping the Bot Fight Mode label. Your third bucket already shows up with the three requests we send: the site refuses the browser while GPTBot and ClaudeBot get the page. The made-up name only matters for telling the first two buckets apart.
On the sites you test, how often does the third bucket come up compared with the first two?
Publishing a null calibration result and keeping the checks with honest weights is more than most tools in this space do. The CDN/WAF number stood out to me. A Disallow line is easy to spot in robots.txt, but a bot challenge at the edge never shows up there, so a site can look wide open and still hand GPTBot a 403. Are you seeing that more from default bot-protection settings, or from rules people added on purpose and forgot about?