← All guides · Published November 14, 2026 · Reviewed November 14, 2026 · By the Ingenious AI team · 6 min read
How Do You Preview a Block Policy Against Your Real Server Log?
The short answer: you paste server-log lines or user-agent strings into AgentBlock's impact preview at /tools/impact-preview, and it classifies the real traffic against your draft policy — explicit decisions if you give them, your built kit if you have one, the default if neither — and reports per agent what the policy would have done. Preview mode only: nothing is applied, nothing deployed. It is how a crawler decision gets priced in your own fetches and citations before it costs anything.
Build my block kit — free · Test a policy in the simulator first
What is the impact preview?
AgentBlock's free tool that replays real evidence — server-log lines or user-agent strings — against a policy and reports per-agent outcomes without applying anything. It runs on the policy you specify, your built kit's policy, or the default policy, in that order.
Why folklore pricing fails
Crawler advice is usually traded in other people's numbers: a publisher's training volume, a platform's citation rates. Your site's numbers are different in the only ways that matter — which agents actually visit, how often, and what the visits are worth in your analytics. A decision priced on folklore is a guess with citations. The impact preview exists so the first input to the biggest crawler decisions is a count of your own last month.
The preview also corrects a systematic bias: owners overestimate famous agents and underestimate quiet ones. The log does not care about fame. anthropic-ai's third of the training crawl, the utility fetcher's daily quiet work, the reported-token noise — all of it appears at its real size, which rearranges most owners' priorities in one reading.
What to feed it
A recent, representative slice: a week of access-log lines is a good start; a month settles arguments. The tool accepts the evidence text directly — log lines or bare user-agent strings — up to a generous size, and classifies each request against the database and your policy. Include a busy period if you can; a preview run on your quietest week will underprice caps and fences alike.
Pair the output with the paths dimension: the preview also takes key paths, so you can see whether an agent's traffic concentrates on content you protect or content you publish. Ten thousand fetches of your image CDN and ten thousand fetches of your paid archive are the same count and opposite decisions; the path split is what separates them.
Reading the three outputs that matter
First, the per-agent refusal list under your draft: every name the policy would refuse, at its real volume. Scan it for dependencies before anything else — this is the friendly-fire check, and the preview is where the preview fetcher's 1,900 monthly visits stop being abstract. Second, the allowed-retrieval volumes, which price the citation channel you are keeping. Third, the UNKNOWN share: traffic the database cannot name, the honest measure of how much of your agent problem name-based policy cannot reach.
Owners often run the preview twice — once on the default policy, once on their intended flips — and diff the two. The diff is the decision, stated in fetches: this many training refusals gained, this many retrieval fetches kept, this much UNKNOWN either way. It is the single most persuasive artifact in a client or partner conversation about crawler policy.
Preview is a mode, not a step
The tool is deliberately inert: it records the preview event for the second-visit picture, applies nothing, deploys nothing. That inertness is a feature owners should insist on in any policy tooling — the right to rehearse against real data without commitment is what makes honest numbers possible, because nobody games a rehearsal. Run it at every policy change, every database version, every quarter's review; the cost is a paste, and the return is that production stops being where surprises live.
When you are ready to act, the sequence continues: simulate the draft on key paths, deploy the kit, verify it by hash, and let the first week's real verdicts confirm what the preview predicted. A policy whose outcomes were previewed, simulated and verified has been examined three times before it can hurt anyone — which is the entire engineering ethic in one workflow.
Five actions to take today
- Feed it a representative slice. A week minimum, a month for decisions; include a busy period so caps and fences are priced at real load.
- Run default and intended policies. The diff between the two runs, in fetches per agent, is the decision stated numerically.
- Scan refusals for dependencies first. Utility fetchers, partners and checkers must be accounted for before the refusal list is approved.
- Read the UNKNOWN share out loud. It sizes what name policy cannot reach and keeps behavioural layers honestly funded.
- Re-run at every change. New database version, new site section, quarterly review — the preview is a paste, not a project.
Worked example: the month that repriced a media site's policy
The Ledger Review, a fictional business-news site, arrived intending to block all retrieval agents on a consultant's advice. Its impact preview on 30 days of logs showed 48,200 training fetches — and 11,700 retrieval fetches, including 4,900 from PerplexityBot concentrated on its explainer library, the section its analytics credited with 2,300 answer-referred sessions a month.
Run against the consultant's draft, the preview priced it precisely: 11,700 retrieval refusals a month to buy zero additional training protection, since training was already fully blocked under the default. The site kept retrieval, blocked training, and fenced its subscriber archive by path.
Subscriber conversions from AI referrals grew 18% the next quarter. The editor's standing rule, taped to the policy: no crawler decision is real until the preview has priced it in our own fetches.
Frequently asked questions
Does the impact preview change my live site?
No. It is preview mode only — nothing is applied and nothing is deployed. It classifies the evidence you paste against a policy and reports the outcomes; your live traffic is untouched, which is what makes it safe to run as often as decisions need it.
What policy does it use if I have several?
Yours, in a defined order: the explicit decisions or blocked IDs in the request if you provide them; otherwise the policy of the kit you have built; otherwise the default policy. That order lets you test hypotheticals without disturbing your built kit.
Is my log data uploaded or kept?
The preview parses the evidence you submit to classify it, and the tool page also offers the evidence tool's offline parser for owners who want log analysis that never leaves the browser. For the preview itself, submit the slice you are comfortable submitting — a week's user-agent lines contain no visitor content.
How much evidence can I paste?
The API accepts substantial submissions — the evidence text is capped at a generous limit sized for weeks of log lines. If your month is enormous, sample it representatively rather than fighting the cap; classification percentages stabilise quickly.
What is a good UNKNOWN share?
There is no universal figure — it depends on your traffic mix. What matters is the trend and the behaviour inside it. A stable, low UNKNOWN share with browser-like patterns is normal; growth, regularity or canary correlation inside UNKNOWN is the signal to strengthen behavioural layers.
Related guides
- How to Block AI Scrapers and Bots from Your Website
- Which AI Crawlers Should You Allow?
- How Do You Block AI Bots on WordPress, Cloudflare, or Plain JS?
- Pre-deploy breakage simulator — test a policy before it ships
- Impact preview — your server log against a draft policy
- How Can You Tell a Spoofed AI Crawler from a Real One?
- Which AI Agents Cannot Be Blocked by Name?
- Why Doesn't AgentBlock Block the Reported GrokBot Token?
- What Can Robots.txt Actually Do About AI Crawlers?
- How Do You Design an Allowlist for AI Agents?
- Block evidence — parse a log, build an evidence pack
- The dated agent database — 25 entries, graded