METHOD 1.0 // SIX ENGINES, 16 CHECKS, TWO PASSES
Methodology
What an LLMsGEO readiness result claims, how it is computed, and how records are held, published and corrected.
What a result claims
Every result on this registry is titled "Readiness for an engine, as assessed by LLMsGEO". It is an assessment of whether that AI engine can reach and read the site, made by the published rules below, version 1.0. It is not a statement made by the engine or by its operator.
- A result never says that an engine lists, favours or mentions a business.
- A result says nothing about visitors or income.
- A result says nothing about the quality or the legal standing of a business.
- A site that has closed itself to a crawler has made its own setting. It is reported as "closed by the site's own setting" or "closed to this engine's crawler", and never as a fault.
Independence
LLMsGEO Registry is independent. It is not affiliated with, and not endorsed by, Google, OpenAI, Microsoft, Perplexity or Anthropic. Their engine and product names are used only to say which engine a result is about.
The six engines and their weights
| Engine | Google AI Search | ChatGPT | Gemini app | Copilot | Perplexity | Claude | Sum |
|---|---|---|---|---|---|---|---|
| Weight | 35 | 30 | 13 | 8 | 7 | 7 | 100 |
The engine weights are a judgment from usage data, not a measurement. They are used only for the overall result. Meta AI, Siri and Grok are monitored and never scored: they have no weight, no check state and no result. Only what robots.txt declares for their crawlers is shown (Meta AI: meta-webindexer; Siri: Applebot and Applebot-Extended; Grok: no testable fact).
The check matrix
Each check has one cell per engine. G is a gate: it carries no points, and a gate with the state fail closes the engine. A number is the points the check carries. S is shown: the state is printed and counts toward nothing. - is not applicable to that engine.
| Check | Scope | Google AI Search | ChatGPT | Gemini app | Copilot | Perplexity | Claude |
|---|---|---|---|---|---|---|---|
| 1 search crawler admitted by robots.txt | on-site | G | G | G | G | G | G |
| 2 user-triggered fetcher admitted by robots.txt | on-site | - | 1 | - | - | S | 3 |
| 3a page directives do not remove key pages | on-site | G | 1 | G | G | S | S |
| 3b page directives do not limit Copilot to title and snippet | on-site | - | - | - | 2 | - | - |
| 4 main content and internal links are in raw HTML | on-site | 1 | 3 | 2 | 1 | 3 | 3 |
| 5 business facts are visible text | on-site | 1 | 3 | 2 | 1 | 2 | 3 |
| 6 delivery hygiene: page size and response time | on-site | 2 | 1 | 1 | 1 | 1 | 1 |
| 7 key pages are in the index the engine depends on | full scope only | G | 1 | G | G | S | 2 |
| 8 Google Business Profile | full scope only | 3 | S | 3 | S | S | S |
| 9 Yelp listing | full scope only | 2 | 3 | 1 | 1 | 3 | S |
| 10 Bing Places listing | full scope only | - | S | - | 2 | - | - |
| 11 Facebook business page | full scope only | 1 | S | S | 1 | S | S |
| 12 trade directories | full scope only | 1 | 1 | 1 | S | S | S |
| 13 review base on the engine's main review source | full scope only | 2 | 2 | 2 | 1 | 2 | S |
| 14 a page that names the service and the city | on-site | 2 | 2 | 2 | 1 | 1 | S |
| 15 pages that answer cost, process and hiring questions | on-site | 2 | 1 | 1 | S | S | S |
| Points, full scope | 17 | 19 | 15 | 11 | 12 | 12 | |
| Points, on-site scope | 8 | 12 | 8 | 6 | 7 | 10 | |
| Applicable checks, full / on-site | 13 / 7 | 12 / 8 | 12 / 7 | 12 / 7 | 7 / 5 | 6 / 5 |
The matrix and the weights are one versioned data set. No check is added, dropped or re-weighted without a new method version and an entry in the change log.
The four check states
Every check, for every engine, ends in exactly one state: pass, fail, not applicable or could not test. The four are never merged, and no page prints a combined figure.
- pass or fail: the rule ran to completion on inputs that were fully read.
- not applicable: the matrix has no cell for that engine, or the rule does not apply to it. The reason is stored.
- could not test: something the check reads was not available. A fetch that was refused, failed or timed out is could not test, never a fail. So is a page our own user agent may not fetch, a response that was cut off, and a record the check needs that nobody supplied. The reason is stored.
Our fetcher being refused says nothing certain about an engine's own crawler, so a network refusal is never a fail. A robots.txt rule or a page directive read from the site is taken as written.
Scopes
The on-site scope is checks 1, 2, 3a, 3b, 4, 5, 6, 14, 15: what the site itself serves. The full scope adds checks 7 to 13, which read business listings and search indexes. Every result names its scope, and no number is carried between scopes.
Only the on-site scope is published. Checks 7 to 13 need listing or index records, and no source for those records exists yet, so they are could not test for every audit. The full scope is still computed and stored with each record, and it is not published.
The on-site rules
- Check 1: search crawler admitted by robots.txt
- Reads the site's robots.txt for each engine's search crawler (Google AI Search: Googlebot; ChatGPT: OAI-SearchBot; Gemini app: Googlebot and Google-Extended; Copilot: bingbot; Perplexity: PerplexityBot; Claude: Claude-SearchBot). Each token is read on its own under RFC 9309: its own group if one names it, otherwise the wildcard group, otherwise no rules. Pass when every token of the engine is allowed on every page read; fail when any is disallowed. A robots.txt that answers "not found" declares no rules and is a pass. A robots.txt that could not be fetched is could not test. The Gemini app needs both of its tokens: when Googlebot is allowed and only Google-Extended is disallowed, the result reads "closed by the site's own setting". This check is a gate for every engine.
- Check 2: user-triggered fetcher admitted by robots.txt
- The same reading for the fetcher an engine sends when a person asks it to open a page (ChatGPT: ChatGPT-User; Perplexity: Perplexity-User; Claude: Claude-User). Not applicable to engines without such a fetcher. The check reports what the file declares; the documentation for ChatGPT-User says robots.txt rules may not apply to it.
- Check 3a: page directives do not remove key pages
- Reads the robots meta element and the X-Robots-Tag header of each page read. Fail when a page carries a directive listed for the engine (Google AI Search: noindex, nosnippet, max-snippet:0; ChatGPT: noindex; Gemini app: noindex; Copilot: noindex, noarchive); otherwise pass. A directive addressed to another crawler by name is not counted. "none" counts as noindex in the two Google columns only. Perplexity and Claude have no directive list: for them the state is not applicable and the declared values are shown as read.
- Check 3b: page directives do not limit Copilot to title and snippet
- As 3a, for Copilot only, with the directives nocache and nosnippet.
- Check 4: main content and internal links are in raw HTML
- Compares each page as fetched with the same page after a browser has rendered it. Pass when at least 80% of rendered words and 80% of internal links are already in the fetched HTML, on every compared page. Both thresholds are a judgment. Page rendering is not enabled on this deployment, so check 4 is always could not test.
- Check 5: business facts are visible text
- Four facts must each appear in the visible text of at least one page read: the business name (compared without case, punctuation or a legal suffix), a phone number, a street address or the locality, and opening hours. Text inside script elements, structured data included, is not visible text. Pass when all four are found; otherwise fail.
- Check 6: delivery hygiene: page size and response time
- Pass when every page read is at most 2,000,000 bytes and was delivered in at most 1,800 ms; fail when any page is outside either line. The time test is whole-response time of our own fetch, a looser test than its time-to-first-byte source. Both are measurements of our own fetch from our own network. The redirect chain of each page is shown and not scored.
- Check 14: a page that names the service and the city
- Pass when a sampled page (the homepage, a service page or a locality page) has, in its title or an h1, both a service term for the business's trade and the name of a Texas incorporated place; fail when no sampled page has both. Service terms are matched as whole words, without case. Service-term lists exist for four trades:
- HVAC: air conditioning, ac repair, a/c repair, hvac, heating, furnace, heat pump, ductwork, air conditioner, cooling
- Plumbing: plumber, plumbing, water heater, drain cleaning, leak repair, sewer, repipe, slab leak
- Electrical: electrician, electrical, panel upgrade, wiring, generator, ev charger, lighting installation
- Roofing: roofer, roofing, roof repair, roof replacement, shingle, metal roof, storm damage, hail damage
- Check 15: pages that answer cost, process and hiring questions
- Reads the headings of the pages read, plus up to 5 answer pages linked from the navigation or the footer. Pass when there are headings about what a job costs, how the work proceeds, and how to choose a provider: one heading for each of the three topics. Fail when a topic has no heading and every selected page was read. When a selected page was not read and a topic has no heading, the state is could not test.
For every page rule: when the submitted page itself was not read, the state is could not test.
Shown, never scored
Some facts are printed as read and carry no points: whether /llms.txt exists, what robots.txt declares for training crawlers, the monitored platforms, the redirect chain of each page, and whether the site uses HTTPS. A training crawler the site has disallowed is printed as "closed by the site's own setting" and changes no result.
The arithmetic
All arithmetic stays inside one engine and one scope. Every division is rounded half up to a whole number. Thresholds are tested on the counts before rounding.
- Applicable checks are the gate and point cells whose state is not "not applicable". Shown cells count toward nothing.
- Coverage is the share of applicable checks that were tested (pass or fail), gates included.
- The 60% rule. When coverage is below 60%, or no check applies, or no check with points was tested, there is no number. The result reads "not enough data". Exactly 60% is enough.
- Gates and the closed state. Otherwise, when a gate check has the state fail, the engine is closed and the result reads "closed to this engine's crawler" (or "closed by the site's own setting", as in check 1) in place of a number. In the on-site scope a closed engine enters the overall result as 0; in the full scope it enters with its off-site points, capped at 25.
- Engine result. Otherwise the result is 100 × the points of the checks that passed ÷ the points of the applicable checks that were tested.
- Overall result. The mean of the engine results, weighted by the engine weights, over the engines that have a result. An engine with "not enough data" is left out of both sums, and the label says how many engines the result rests on. The overall result is withheld, with no number, unless Google AI Search and ChatGPT both have a result.
- Ten-point view. A result divided by 10, rounded half up, and never below 1. It is given only beside a result that is a number.
Every number is printed with its label: the title, the scope, how many applicable checks passed, how many were not applicable, how many could not be tested, the coverage, the scan time, the method version, the trade profile and the pages read.
Two passes
Every audit is two complete scan passes. The second pass starts at least 60 minutes after the first ends, and within 24 hours. When both passes return the same check states the audit is final, and the first pass is the record. When any state differs the audit is held: nothing from it is published, and the record is left out of every public page, list, count and feed until a later audit is final. An audit whose second pass was not run within 24 hours is held as well. A held record is scanned again after 1, 3 and 7 days.
Check 6 reads a measured time, so a site close to its line can differ between passes. That is the rule working as designed.
What is stored
For each pass the registry stores one row per check per engine (the matrix cell, the state, the reason or an evidence reference), the fetch record of each request (status, size, time and redirect chain) and the SHA-256 hash of each response body.
Every pass is also stored as evidence: the raw response bodies, the response headers and a manifest with a SHA-256 of each object, in locked object storage where a stored object cannot be overwritten. Any stored pass can be re-scored from those snapshots alone. The public pages print the check states and the fetch facts; they do not republish the stored bodies.
The scanner
The scanner fetches with plain HTTP requests and runs no page script. Its user agent is:
LLMsGEO-Engine/1.0 (+https://llmsgeo.com/methodology)
It reads robots.txt before anything else and obeys the rules addressed to LLMsGEO-Engine, or to the wildcard group when none names it. To keep the scanner off a site, add this to the site's robots.txt:
User-agent: LLMsGEO-Engine Disallow: /
The pages it may not fetch are then could not test, and never a fail.
One pass sends at most 15 requests, waits 1,000 ms between requests, follows at most 5 redirects per page (each one checked again), reads at most 5 MiB per response, waits at most 20 seconds for one response and stops after 75 seconds. These limits are operator settings, not sourced values. Private, loopback, link-local and reserved network addresses, embedded credentials and non-default ports are always refused, at submission and at every redirect. The scanner never follows the submitted page to a different host.
Records reach the registry in three ways, all scanned by the same method and all given a permanent scorecard address and the same 14-day hold:
- Authenticated submissions. Verified submitters queue a site through a private endpoint that requires a secret key. The endpoint validates the submission and never fetches the site; the scheduled scanner does.
- Public checks. Anyone can run a first pass on a public site at /scan. A public check takes a domain name, not a bare IP address, and is limited to one check per domain per hour, a number of checks per network address per hour and a number of checks overall per hour. Its record is unlisted: it has a scorecard address, is never shown in the public directory and stays noindex. A public check never overwrites a record created by an authenticated submission and never changes an existing record's name or location.
- Scheduled scans. The scheduled scanner runs one pass at a time. A second pass that is due comes before a queued first pass.
Out of date after 90 days
Each label stores the time 90 days after its scan. Past that time the pages print "out of date" beside every number from that audit, and the numbers are no longer shown as current.
Mandatory 14-day hold
Every new record is held for 14 days, with zero exceptions. The hold runs from a site's first scan and its release date is fixed at that moment: later re-scans update the scorecard at the same address but never move the release date. During the hold the public directory and the scorecard show only "Pending" and the release date: no number, no engine result and no check state is shown anywhere until then. The scorecard address is withheld from the public directory and feed until the release date. The rule is enforced in the database itself, so no request can shorten or extend it.
Each scorecard lives at a permanent, unguessable address made of the domain plus a random 48-bit token. The address never changes when a site is re-scanned. Scorecards inside the hold are marked noindex so search engines do not list them before release, and unlisted scorecards stay noindex.
Free owner corrections and re-checks
Site owners may change their site at any time, with no involvement from the registry. Owners may request a free re-check before publication, a factual correction, or a removal review through the corrections form. Requests are free, need no account and collect no contact details. The form accepts a limited number of requests per network address per hour.
The public directory
The homepage lists the listed records in alphabetical order by business name, with trade and city. A record inside its hold is marked "Pending" with its release date and has no scorecard link until then. Every page is served from this site alone: no map tiles, scripts or other files are loaded from anywhere else.
Neutral operation
The registry is vendor-neutral, carries no advertising, and does not direct site owners to any service provider. It publishes no email address: the corrections form is the only way to reach it.
Change log
- 2026-11-17
- Publication only; no check, weight or arithmetic changed and the method stays version 1.0. Scorecards inside the 14-day hold show "Pending" and the release date only. A record whose two passes disagree is left out of every public page until a later audit is final.
- 2026-11-15
- Check 14 activated with service-term lists for four trades and the Census list of Texas places; every scan pass stored as evidence; engines with fewer than 4 tested scored checks marked limited data.
- 2026-11-15
- Method version 1.0. One readiness result per engine, from 16 checks and two scan passes per audit. It replaces the earlier fixed-point standard. Every result produced under that standard was withdrawn.