A visibility reading is a sample of answers taken on one day. That sentence contains three reasons the number will move.
The prompt set
Brightwater Dental, a fictional client, twelve questions frozen on 2026-09-14.
1. best dentist in brightwater
2. emergency dentist near brightwater
3. how much does a dental implant cost
4. are dental implants painful
5. how often should you get a check up
6. dentist open on saturday brightwater
7. is teeth whitening safe
8. how long do veneers last
9. nhs dentist waiting list brightwater
10. childrens dentist brightwater
11. how to stop bleeding gums
12. dentist that does payment plans brightwater
Six branded, six unbranded. That split matters, because an assistant that names a business for its own name is not a discovery.
Two readings, fourteen days apart
reading A, 2026-09-14
prompts run: 12
answers collected: 12
site named: 5 of 12 (prompts 1, 2, 6, 10, 12)
pages cited: /emergency, /implants, /childrens
reading B, 2026-09-28
prompts run: 12
answers collected: 12
site named: 7 of 12 (adds prompts 3 and 9)
pages cited: /emergency, /implants, /childrens, /pricing
difference: 2 answers. Both are unbranded cost questions.
Two answers out of twelve is a difference of seventeen points. That is still inside the noise of a small sample.
What can explain the difference
- the assistant changed model or behaviour between runs
- one answer quoted a page that was edited in between
- personalisation showed a different answer than the one recorded
None of those is a change you made. Say so before somebody reads the two numbers as a trend.
The reporting block, ready to copy
measure: mentions in sampled answers
prompts: 12, frozen on 2026-09-14, version 1
run on: 2026-09-28
mentions: 7 of 12, up from 5 on 2026-09-14
cited pages: /emergency, /implants, /childrens, /pricing
type: estimate
limits: small sample, model version not recorded, no customer
behaviour measured
next: read the four cited pages against the seven prompts
that did not mention the site
Troubleshooting
| Symptom | Likely cause | What to do |
|---|---|---|
| Mentions fall to zero | The prompt set or the model changed | Version the set, and re-run it unchanged |
| Every branded prompt mentions us | That is the easy half | Report branded and unbranded separately |
| An answer quotes a page we deleted | The index is stale | Check the page, then re-run |
| A client asks for a citation guarantee | Nobody can promise one | Report the sample and the method |
| Two runs disagree wildly | Sample too small | Add prompts before adding conclusions |
Acceptance: the reading is usable when
- the prompt set is frozen and versioned
- branded and unbranded prompts are counted apart
- the run date and the number of answers are recorded
- cited pages are listed, not summarised
- the report says estimate, and names the limits
Limits
Visibility is not demand, and a citation is not a customer. Keep the raw answers, so the next reading can be compared honestly rather than remembered conveniently.
One last habit is worth keeping. Write the reading into the client’s record beside the pages it cited. The next conversation then starts from the answers, not from a memory of them.