Monthly. A law firm should publish one AI visibility number each month, built from the same frozen prompt set, and treat anything more frequent as raw material to pool. A weekly run on a small sample moves up and down by chance, and the fixes a firm makes take weeks to be re-read by the engines.
Firm AEO's tracking post gave this question one FAQ answer and its timeline post set out the five clocks. This post does the cadence arithmetic itself: how much a weekly number wobbles by chance, what the engines' own documentation says about re-reading a page, and the three events that justify an unscheduled check.
- A law firm should report one AI visibility number per month, built from a fixed prompt set, and treat weekly runs as inputs that get pooled rather than as separate results.
- At a true appearance rate near 30%, a 15-answer weekly sample carries a 95% interval of roughly plus or minus 23 points, so most week-to-week movement is sampling noise.
- Google's Search Central documentation says crawling a new or changed page can take a few days to a few weeks, so a page edited on Monday has usually not been re-read by Friday.
- Three events justify an extra check outside the monthly run: a robots.txt change, a page rewrite that has been indexed, and a major engine model release.
Why is a weekly AI visibility number mostly noise?
Because a weekly run on one market produces too few answers for the percentage to hold still. Share of answers is a count divided by a count, so with a small sample the figure swings from chance alone. Citorian's June 2026 test logged 359 answers from 375 query runs across five metros and five engines, and a single firm in a single market works at a fraction of that scale.
The table below applies the standard error of a proportion to a firm whose true appearance rate is 30%. The 30% is illustrative, chosen only to show the arithmetic, and the intervals are approximate. The point is the shape: quadrupling the sample halves the swing.
| Sample | Answers | Typical swing by chance (95%) |
|---|---|---|
| One weekly run, 5 prompts x 1 engine x 3 runs | 15 | About plus or minus 23 points |
| Four weekly runs pooled into one month | 60 | About plus or minus 12 points |
| Monthly, 10 prompts x 4 engines x 3 runs | 120 | About plus or minus 8 points |
| Quarterly pool of that monthly set | 360 | About plus or minus 5 points |
How fast do the engines re-read a page a firm just changed?
Slower than a week, in the documented cases. Google's Search Central documentation says crawling a new or changed page can take anywhere from a few days to a few weeks, and that requesting a recrawl more than once does not speed it up. A practice page edited on a Monday has usually not been re-read by that Friday, so a Friday measurement is testing the old page.
OpenAI documents one faster path. Its crawler documentation says it can take about 24 hours after a robots.txt update for ChatGPT's systems to adjust. Access is therefore the one lever whose effect can show up inside a week, and it is the first of the three exceptions covered below.
Can a firm run weekly checks and still report monthly?
Yes, if the weekly runs are pooled before anyone reads them. Four weekly runs of the same prompt set, scored the same way, add up to one monthly sample four times larger than any single week. The monthly number is the sum of the weekly counts, and the individual weeks stay in the ledger for diagnosis only.
What does not work is reading each week as its own result. A firm that celebrates a rise from 27% to 40% between two 15-answer weeks, or alarms over a fall, is reacting to a swing that the table above shows is within chance. Firm AEO keeps weekly figures out of client reports for that reason.
Which events justify a check outside the monthly run?
Three events change the underlying facts quickly enough to be worth a targeted look: a crawler access change, a page rewrite that has been indexed, and a new model release from an engine the firm measures. Each check is narrow. It asks one question about one change, and it is logged beside the monthly number rather than folded into it.
- After a robots.txt or firewall change: re-test within a few days, because OpenAI documents about 24 hours for its systems to adjust. Confirm that OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot and Bingbot can fetch the page.
- After a rewritten page is indexed in Google and Bing: check that the URL appears in the cited-sources column for the prompts it targets. Count it as a diagnostic, not as a share-of-answers data point.
- After an engine ships a new model or changes its search behaviour: re-run the full frozen set once, label the run with the model change, and keep it out of the monthly trend line.
What should be measured quarterly instead?
The slow clocks. Directory selections, review accumulation and the weaker engines move on a timescale of months, so reading them every month adds cost without adding information. Super Lawyers selects attorneys once a year on a state-by-state basis, and the iLawyerMarketing survey of 1,110 US adults published on August 17, 2026 found 6.2% would use Claude and 3.7% Perplexity for lawyer research, against 41.9% for ChatGPT.
Firm AEO therefore runs the main engines monthly and the smaller ones quarterly, and reviews directory profiles on the same quarterly rhythm. The schedule below is a typical starting point, not a rule, and a firm in a fast-moving market may add a check.
| What | How often | Why |
|---|---|---|
| Share of answers, main engines | Monthly, weeks pooled | Large enough sample to read |
| Cited-sources list and domain counts | Monthly | Shows which pages and directories the engines read |
| Smaller engines (Claude, Perplexity) | Quarterly | Low consumer reach per the 2026 iLawyerMarketing survey |
| Directory profiles and review recency | Quarterly | Slow clocks; Super Lawyers selects annually |
| Robots.txt and crawler access | On every change | OpenAI documents about 24 hours |
What can a firm say publicly about a monthly number?
It can state the number with its method attached. ABA Model Rule 7.1 prohibits false or misleading communications about a lawyer's services, and a percentage published without the prompt set, engine list, run count, sample size and date can mislead. A cadence also affects what is honest to say: a figure from one weekly run is a sample of 15, and calling it the firm's AI visibility overstates what was measured. Firm AEO is a marketing company, not a law firm, and the firm's ethics counsel decides what the page says.
Is a weekly check ever worth it?
Yes, as an input. Four weekly runs of the same frozen prompt set pool into one monthly sample, which is large enough to read. A single week should not be reported as the firm's visibility.
Should a firm check after every blog post it publishes?
Only the cited-sources column, and only after the page is indexed in Google and Bing. Google says crawling can take a few days to a few weeks, so a check the same week usually tests the page before it has been read.
Does a bigger prompt set let a firm measure more often?
Partly. A larger sample narrows the swing, but the engines still re-read pages on their own schedule, so a faster cadence cannot see a change that has not happened yet. Sample size and re-read speed are separate limits.
What if the monthly number drops?
Check the ledger before acting. Confirm the prompt set was unchanged, review the failed-run count, and compare the drop with the interval for that sample size. A fall inside the interval is usually noise, and a fall that holds for a second month deserves investigation.
- 1.Citorian, which personal injury lawyers AI recommends (June 2026)
- 2.Google Search Central, ask Google to recrawl your URLs
- 3.OpenAI, overview of OpenAI crawlers
- 4.iLawyerMarketing, What online sources do people use to research and find attorneys in 2026 (August 17, 2026)
- 5.Super Lawyers, selection process
- 6.ABA Model Rule 7.1, Communications Concerning a Lawyer's Services