How we measure Site Pulse
Site Pulse crawls your site and runs 22 modules over the pages it fetched. Every finding is a deterministic read of bytes already in hand, with no model in the path.
Depth: crawls per scan
Would a number hold still if we ran it again? On this audit it would, so there is one rung and nothing to buy.
| Crawls per scan | Grade | Cost | What it buys |
|---|---|---|---|
| - | T1Snapshot | - | Not needed. One pass is already Verified. |
| - | T2Directional | - | Not needed. One pass is already Verified. |
| - | T3Consensus | - | Not needed. One pass is already Verified. |
| - | T4Established | - | Not needed. One pass is already Verified. |
| 1 Default | T5VerifiedChecked directly. Same answer every time. | ×1 | Repeat runs agreed on 2,817 of 2,821 checks. |
The grade is unstarred because we ran the validation, joining every stored scan to the one before it on the same template, check by check. The three pairs re-run within six hours with no rule change between them agreed on all 961 shared checks, and the checks that reach the network (AI-bot probes, identity links, the certificate read, page weight) on all 408 of theirs.
A day later things do move, by about eight per cent at seventeen hours and sixteen at thirty-six. That is your site changing, which is what snapshots are for. Sampling would also multiply the crawl that lands on your server: we cut the deepest preset from 250 pages to 150 in September 2026 after a scan overwhelmed a customer on shared hosting.
Breadth: pages
How much of your site we saw. The rungs are the template's three crawl depths, charged at 10, 30 and 80 credits.
| Pages | Grade | Cost | What it buys |
|---|---|---|---|
| 1 | UnknownNot recorded, or too few samples to tell. | ×1 | Quick. Home page only. |
| - | T1 | - | Not on the scale. Narrow is the lowest breadth grade. |
| - | T2Narrow | - | Not offered. The first rung already reaches T3. |
| 50 Default | T3PartialStill finding new things at the end. | ×3 | Standard. |
| - | T4Mostly saturated | - | Not offered. The ladder steps from T3 straight to T5. |
| 150 | T5Saturated*The last samples found nothing new. | ×8 | Full. |
* Our estimate, not yet validated by real runs.
The 50-page rung promises less than most runs at it measure, on purpose: the sites still turning up new page shapes at page fifty are the ones where the answer matters. The one-page rung is unstarred because a single page says nothing about the pages beside it.
What the run itself measures
Every scan stamps its own grade from its novelty curve. The unit is a page shape: the set of checks that warned or failed on a page, taken as one key. A clean page is a real shape too. We tried counting new kinds of issue first and dropped it: the list of checks is closed, the last new kind appeared by page seven in every stored 50-page scan, and every site graded saturated. Page shapes spread properly - one scan found three across fifty pages, another was still finding new ones at page forty-two.
A crawl that follows every link and sitemap entry and runs out of site before its page budget is a census, graded saturated; reading the curve instead would have told 51 of 77 finished crawls that most of their site was unseen. A page nothing links to and no sitemap mentions stays invisible to any crawler, ours included. One-page scans, and every scan run before this measurement existed, report breadth as unmeasured.
What the report claims
| Number | Type | What it can support |
|---|---|---|
| Every finding | Verified | A title is present or not, a canonical points somewhere or not, robots.txt names a crawler or not. A count is never printed as a rate: eighteen failing checks is eighteen findings. |
| Category and snapshot scores | Derived | The weighted share of checks that passed, over every check that passed, warned or failed. Information findings sit outside the denominator; failures stay in. The snapshot is the mean of categories that produced findings, skipping empty ones rather than scoring them zero. |
| The Signal narrative | Interpreted | Its severity is computed from the score and failure count by a fixed rule; only the wording is a model's, a single read over final findings. It moves no number. |
Fidelity
We read the real thing: every check works on the response your own server returned to our crawler, GeorgeBot, at one request a second or one every five when a template asks us to tread lightly. Nothing comes from a third-party index or a cached copy.
- Unless a template turns on rendering, we read HTML as it arrives rather than as a browser assembles it - exactly what an AI crawler receives - and the LLM view module fails any page where almost no text survives.
- The live AI-bot probes leave from our own address, not an operator's, so a pass rules out a user-agent block and nothing more. They never grade worse than a warning for that reason.
- A crawl cut short by a daily request budget carries a finding naming the pages fetched and the limit.
What we do not claim
- That your site has no other problems - only that these checks, over these pages, returned these answers.
- To have seen pages nothing links to.
- A confidence percentage on the Signal or anything else a model wrote.
- That a passed AI-bot probe means the bot can reach you, only that your server did not turn away something claiming to be it.
- A score for a category with no findings. It scores nothing rather than zero, because we did not measure it.
