Three pages. Three different evidence problems.
This example shows how SeeLLM separates what happened from what it might mean—then turns that boundary into a reversible action and a verification window.
- Example subject
- Northstar Analytics
- northstar.example
- Observation window
- 2026-07-01 — 2026-07-31
- Pages reviewed
- 3
Read before using
This is an illustrative report for a fictional company. The evidence patterns are realistic, but the domain, counts, samples, and outcomes are not customer results.
Executive read
Fix access first. Improve source fitness second.
Fix crawler access to installation docs first, then strengthen the pricing and comparison pages for the sampled questions where other sources were preferred.
01
/docs/installSearch crawler blocked
What we observed
Access response
A request using the Claude-SearchBot identity received HTTP 403 while a standard browser request received HTTP 200.
Method: Public user-agent access check
Likely access gap
The differing responses indicate that an edge or bot rule is the most likely reason this crawler cannot retrieve the page.
Method: Comparison of crawler and browser access responses
What it means
The installation page is unavailable to the tested search crawler, so content improvements on this URL cannot help that crawler until access is restored.
What it does not prove
This single access check does not prove that every Claude product, crawler IP, geography, or request path is blocked.
Recommended action
Restore search-crawler access before rewriting the page
Review the edge rule producing the 403, preview a narrow allow rule for search and user-directed retrieval identities, and leave training-crawler consent as a separate decision.
- 1.Identify the Cloudflare or origin rule matching Claude-SearchBot.
- 2.Preview an allow rule scoped to the public documentation path.
- 3.Keep authenticated, admin, and private documentation excluded.
How we will verify
Claude-SearchBot access check returned HTTP 403 on 2026-07-31.
- Claude-SearchBot receives HTTP 200 or a canonical redirect followed by HTTP 200.
- Browser access remains HTTP 200.
- No private or authenticated path becomes public.
Repeat immediately after the change and monitor for 14 days.
02
/pricingCrawled, absent from citation sample
What we observed
Crawler visits
OAI-SearchBot requested the pricing page 18 times during the observation window.
Method: Server-side request log
Citation sample
The pricing URL appeared in 0 of 12 recorded answers across a fixed set of plan-fit and pricing questions.
Method: Recorded answer sample across ChatGPT search and Perplexity
Crawled, not cited
Access exists, but the page was absent from this citation sample; plan facts and comparison context are the next things to inspect.
Method: Relationship between observed requests and sampled answers
What it means
The page is reachable and repeatedly requested, but it did not appear as a source in the defined answer sample.
What it does not prove
The sample does not prove that the pricing page is never cited, ignored by every model, or responsible for the absence by itself.
Recommended action
Make plan facts easier to extract and compare
Publish a concise plan-fit table, keep pricing and limits consistent with machine-readable sources, and answer the sampled buyer questions directly on the page.
- 1.Add a plan-fit table with audience, price, limits, and exclusions.
- 2.Link the visible plans to the canonical machine-readable pricing file.
- 3.Add direct answers for the recorded pricing and plan-fit questions.
How we will verify
0 of 12 recorded answers cited the page in the July sample.
- Pricing facts remain consistent across HTML, JSON-LD, and pricing.md.
- The page appears in at least one repeated answer sample using the same questions.
- Crawler requests continue without a new access regression.
Repeat the same answer sample weekly for 28 days.
03
/compare/warehouse-toolsCompetitor source gap
What we observed
Source gap
A competing comparison source appeared in 7 of 12 recorded answers; Northstar's comparison page appeared once.
Method: Recorded source comparison using the same question set
Likely content gap
The competing source answers deployment, data-retention, and pricing questions that Northstar's page leaves implicit.
Method: Page-content comparison against sources in the recorded sample
What it means
Another source more consistently answered the sampled evaluation questions, giving Northstar a concrete comparison-content gap to address.
What it does not prove
Source frequency in this sample does not establish general market preference, answer quality, or future citation behavior.
Recommended action
Answer the comparison questions the preferred source covers
Add a balanced, source-backed table for deployment, retention, pricing, and best-fit criteria, including cases where the alternative is the better choice.
- 1.Add explicit comparison criteria tied to the recorded questions.
- 2.Cite primary sources for competitor facts and show the checked date.
- 3.State who should choose each option instead of declaring one universal winner.
How we will verify
Northstar appeared once and the competing source appeared 7 times in 12 answers.
- Every comparison claim links to a primary source and checked date.
- The repeated sample shows whether the source gap narrowed.
- The page remains balanced when the alternative is a better fit.
Repeat the fixed question set weekly for 28 days.
Method and limits
Full evidence model- —The company, values, answer samples, and outcomes in this example are illustrative.
- —Crawler identities can be spoofed unless vendor network ownership is verified.
- —Generated answers vary by model, prompt, geography, account, and time.
- —A missing observation is not proof that no activity occurred.
Your pages, next
Start with access. Earn the right to infer.
Run the public preview on an important page, then connect the Site Agent when you need observed requests, referral evidence, and before-and-after verification.