Summary
- The bot-score page prints a 1-to-99 scale and five categories; the four numeric rows tile that range exactly (1 + 28 + 70 = 99), and one integer separates the single value called automated from the band called likely automated.
- The same page says model updates do not change your rules while bot scores can shift as the model learns, then tells the customer to periodically review those rules — a duty with no interval, no owner and no event that fires.
- Score 0 is defined as "not computed" and sits outside the stated 1-to-99 range; the page prints no guidance on how a rule should treat it, and this article quotes no field name because the reference page that would print one was not obtained.
The definition is a single sentence: "A bot score is a score from 1 to 99 that indicates how likely that request came from a bot." Ninety-nine values, one per request, issued by the network that sees the request rather than by the site that receives it.
The table does arithmetic nobody had to write
Under the definition sits a five-row category table. Four rows carry numbers: not computed, 0; automated, 1; likely automated, 2 through 29; likely human, 30 through 99. The fifth, verified bot, carries a description and no range at all — "non-malicious automated traffic (used to power search engines and other applications)". One category in five is therefore not a number on this page; it is a status.
The four numeric rows tile the stated range with nothing left over: 1 + 28 + 70 = 99. That is my addition, not a figure Cloudflare prints, and it matters because of where the seams fall. The one value the page calls automated is 1. The band it calls likely automated begins at 2. A single integer separates a verdict from a suspicion, so the sharpest boundary in the whole instrument sits at the bottom of the scale, where a threshold written as "less than three" would wipe out the entire automated band and one value that means not computed.
The likely-human floor is 30. On a scale whose midpoint is 50, the lowest value the documentation calls human still sits twenty points down, and the likely-human band covers 70 of the 99 values — about 70.7 per cent on my division. Suspicion is a narrow strip at the bottom. Reassurance is most of the dial.
The name points the wrong way
A bot score sounds like a measure of botness. It is not. The page states that the engine uses supervised machine learning, and that the output variable is "the predicted probability that a client is human, mapped to the final 1–99 bot score". The model predicts a person; the score reports the result on a scale that the product named after the opposite.
Once you know the polarity it is coherent, even elegant. But a rule author who reads the field name and assumes the big numbers are the dangerous ones has built a filter that blocks readers and admits scrapers, and no amount of documentation polish rescues an expression written by someone who skipped the sentence. The page says which way the scale runs. It cannot make anyone read it.
The sentence that turns a metric into a chore
Here is the pair that makes this a management problem rather than a measurement one, quoted as printed: "Model updates do not change your WAF rules, challenge flows, or Bot Management configuration. However, bot scores can shift as the model learns, so periodically review rules that use bot score to confirm they still work as intended."
The first sentence is a compatibility promise: your configuration survives the release. The second corrects what surviving means, because the rule survives and the quantity it compares against does not. Nothing in the second contradicts the first, which is exactly why it is easy to skim, and why a team can honestly believe it has a stable control when it has a stable expression over an unstable input.
And "periodically" is not a period. No interval, no owner, no record, no notification when a rule starts catching a different population than the one it was tested on. The vendor revises the measurement; the customer owns the discovery. That is not a bug in the wording — it is the allocation, printed plainly, and it is the most consequential sentence on the page.
The features are pooled
One more line earns its weight. The inputs are "request features, including headers, session characteristics, and browser signals, collected from traffic across the Cloudflare network." Not the customer's traffic alone. The evidence base is the network's.
For a large property that is a comfort: far more traffic than any single site could contribute has informed the model. For a small one the shape is stranger, because the statistics that decide a request to a fifty-visitor-a-day site were learned mostly somewhere else. No page in the material read prints a per-feature weight, so a customer cannot say which of those outside patterns moved its own verdict on any particular request.
What is published, and what merely is not
The page names the model class, the input families and the output variable. It prints no coefficient, no per-feature contribution, no training window, no accuracy figure. It also prints no sentence saying that any of this is withheld on purpose. Those are separate findings and only the first is documented: the machine's shape is described and none of its numbers are published.
The logging line is genuine but narrow. Customers "can analyze the request features this model uses in your own logs, through Logpull or Logpush" — the features. Not the weights, and not an explanation of why one request landed at 27 and another at 34.
What the visitor receives
Nothing. A reader whose browser traffic is scored low and blocked at the edge sees an interstitial or an error — no classification, no band, no reason, and no party to ask in the page itself. The customer can log features. The visitor cannot request the file.
Whether any route exists for the operator of scored traffic, I am not going to assert. The verification reference I reached came back as an index naming "Web Bot Auth" and "IP validation" without stating criteria on itself, so how a firm's traffic becomes verified is not described here, and neither is the absence of a review process I never reached to. An unread page is a limit on me, not a finding about Cloudflare.
What I read, and what it does not cover
One page: the bot-score concept page in Cloudflare's developer documentation, read on 17 November 2026. The WAF custom-rules field reference — which would carry the literal field name, its type and any plan gating — was not obtained in this pass; the page returned was the custom-rules overview, which mentions bot score as a navigation link and prints no field row. So this article quotes no field name and makes no claim about comparison syntax. It gives no figure for how often a review should happen, how many requests are scored, how many are blocked, what accuracy the model reaches, or which plans include the feature.
The thing worth writing down
Two facts share one page and neither cancels the other: Cloudflare publishes the scale and Cloudflare reserves the schedule. A customer who writes a threshold has made a promise to visitors about what will not reach them, and the vendor's own advice is to come back later and confirm the promise still holds.
The scale is documented. The clock is not. Any team running a bot-score rule in production that cannot name the date it was last reviewed is operating a control it inherited rather than one it chose — and the documentation already told them, in plain print, that this is how the arrangement works.
Cloudflare Inc in the BTW directory holds the entity record behind this briefing.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






