How we measure AI visibility

Weekly runs use the approved set and frozen roster. Actual collection timestamps are recorded separately from the reporting week.

Weeks use ISO boundaries in UTC. Plan allowances are on the pricing page; confirmed question sets may be smaller.

Prepaid checks run only when requested. Packs do not expire; each check uses one credit on a business you own.

One extra check a business a day: press again the same day and you are taken to the check that is already running, and nothing comes off your balance.

The prepaid shape is 5 questions across 3 assistants, once each: up to 15 answers. It does not inherit a larger plan’s weekly roster.

Extra checks keep their own answers, date and count. They are excluded from weekly headlines, weekly email and monthly measurement totals.

Changing the weekly question set or roster affects future runs and may break direct comparisons.

Optional add-ons are monthly, prorated for the remaining period, and end with the plan. Add or remove them from billing; none is selected by default.

Add-ons affect future planned runs. Earlier questions, rosters and results stay frozen; a changed design is not directly comparable.

Grok (on Starter)

10 answers more a week on top of what the plan already asks, which is about $1.10 a week of what we pay the vendors for them.

Claude (on Starter)

10 answers more a week on top of what the plan already asks, which is about $0.24 a week of what we pay the vendors for them.

Ten more questions (on Starter)

90 answers more a week on top of what the plan already asks, which is about $0.68 a week of what we pay the vendors for them.

One more business

One more business on the account, beyond the ones a plan already allows. It asks nobody anything and costs us nothing to run: the business in it pays for its own checks through its own plan, and a business on Free is sampled by nobody.

Ten more questions (on Growth)

110 answers more a week on top of what the plan already asks, which is about $2.01 a week of what we pay the vendors for them.

Ten more keyword reads (on Starter)

10 more phrases read in Google every Monday, which asks no assistant anything and buys nothing from one: it is about $0.52 a month of what we pay the search data provider to read them.

These vendor-cost estimates come from the configured cost reference. Each actual request records its own cost.

Unused prepaid checks are refundable within 14 days. Unlike monthly add-ons, credits do not expire.

We report whether a saved answer named your business. That is not a stable position in an AI-generated list.

Inspect the exact question, response and cited sources behind each count.

Tracked phrases use DataForSEO SERP observations, to 50 results on desktop, in your market. Each observation records its date and provider.

A position in an ordered search result differs from a mention in an AI answer. Neither explains or guarantees the other.

A keyword reading is one observation, not a confidence interval. Successive readings are shown as facts without an inferred direction.

“Not in the first 50” differs from a failed lookup. We cannot distinguish lower results from absence.

The AI Overview flag records whether an overview appeared; it does not claim that the overview named you.

Starter allows 10 tracked phrases, Growth 50. Stopping frees the slot and keeps historical readings; restarting resumes that history.

Wilson intervals assume independent binary observations with a common mention probability. Repeated answers can be correlated, so the bounds are approximate.

Intervals describe sampling uncertainty. They do not forecast future results or prove that a fix caused a change.

Measured intervals and fixed-rate calculator illustrations are separate. Illustrations assume a rate and do not represent collected answers.

More questions or a different roster changes the measurement design; it does not merely sharpen the existing result.

We have not validated empirical interval coverage on our own samples. The method is not claimed to be verified in practice.

The table identifies enabled assistant APIs and vendor SERP sources. Plan-only providers are included only in rosters that carry them. Historical runs keep their original coverage.

A separate name extractor reads the final answer and stores the business names it finds. Retrieved pages alone are not cited or named-answer evidence.

A name in one response establishes only that it appeared there. It is not an endorsement, stable position or comparison of business quality.

Use the full panel and its coverage. One favourable response is not evidence of market-wide performance.

Each diagnostic labels its source: fetched evidence, vendor observation, or a cited research threshold.

Google Maps listing (vendor)

A vendor listing matched to your business and location, with match confidence. It does not verify completeness or accuracy.

Google reviews (research)

Listing reviews compared with a cited research threshold. This is not a measurement of business quality.

Local best of pages (vendor, then measured)

Relevant roundup pages and whether they name you. We fetch at most three when robots permits; editorial reasons are unknown.

AI crawler access (measured)

Robots rules for tested public paths. Permission is not evidence of visits; unreadable or truncated rules remain unknown.

Bing index (vendor)

Results from a public site query. It is neither a Google index check nor private Bing Webmaster data.

Business details in your site (measured)

Recognized business or product identity in homepage structured data. Details on other pages may not be seen.

Unreadable evidence is unknown, not a pass. Missing website-dependent checks are listed explicitly.

Changes to your site or profile require your approval, including explicit standing authorization where enabled.

A draft is a proposal. An approved change records the exact target, before/after, actor, time and outcome; undo creates a linked record.

Standing publication is a separate explicit authorization. It retains per-page records and a hold during which you can edit or stop publication.

Each draft answers one approved question using supported business facts.

The plan cap is 4 pages per calendar month. A lower configured frequency can reduce that cap.

Supported inputs are confirmed profile facts, saved answers naming you, your cited website pages, and permitted review quotations. Unsupported sentences or figures are refused; sources appear with the draft.

Published pages disclose how they were drafted and who approved them.

Approve each page or explicitly authorize standing publication. Standing pages keep their own approval record and configurable hold; ending the standing leaves published pages intact.

Before/after counts describe observations only. Method changes can break comparisons; interval separation does not establish cause.

Google’s published content policies remain relevant:

Google does not refuse a page because a model helped write it. What it refuses is many pages made for a machine instead of for a reader. Its own guidance on content written that way asks for two things, and both of them are in the paragraphs above.

A partners page lists up to 5 reciprocal recommendations, with the business name, trade, location, owner-written note and pairing month.

Both businesses must approve. Each action leaves its own record; pairs are not assigned automatically.

Pending asks count against both caps. Same-trade businesses, indistinguishable trades, one site, or businesses in one company’s workspace cannot pair.

Google’s published link policies remain relevant:

A reciprocal page remains reciprocal even with a cap and disclosure. Link qualification does not guarantee policy compliance.

Partner links served by PicksYou use rel="nofollow". Customers choose the recommendations; paying customers are not editorial endorsements.

Either party can end a pair. Missing links for 14 days or two weekly checks end it; unanswered asks expire after 14 days. Ending a partners page or deleting an account also ends its pairs. Undo records are written on both sides.

Changes appear immediately on PicksYou and on the next plugin refresh on the customer site. Hourly caching is required for an hourly expectation; we verify removal.

No partner link guarantees a mention. Any before/after reading is observational and may be incomparable.

9 spreadsheet exports use UTF-8 with a byte-order mark. Each states its row grain and preserves recorded dates.

The weeks

One row per sampled reporting week: named count, denominator and available bounds. Missing intervals stay empty.

One week's answers

One row per answer: provider, exact question, named verdict and UTC collection time.

What was in the way

One row per diagnostic: date, recorded evidence and source type.

Who arrived

Recorded visits by day and assistant. Missing telemetry does not prove zero visits.

Named beside you

Named businesses by week, with their count and denominator. Growth exports the history; Starter shows it in the account.

A closed month

Frozen reporting weeks and provider totals. Open months have no closed-month file.

What you approved

Approval target, actor, outcome and verification. Undo is separate; full before/after content stays in the approval record/account export.

Your keywords

Phrase, observed position, found page, overview flag, time, provider, market, depth and device. Failed readings differ from absent results; stopped phrases retain history.

The sites they cited

Cited hosts, citing-answer counts and denominators. The export includes all hosts; retrieved-only sources do not count as citations.

Spreadsheet exports do not include full assistant prose or individual cited-page lists. Saved answers remain accessible in the account.

The API exposes measurements and authorized actions; MCP tools are read-only. Public proof links can be revoked.

A workspace holds businesses and invited members. Access is scoped to selected clients or an explicitly granted all-client scope; people are not billed as seats.

Viewer: read and export. Admin: run work. Owner: manage money. The permission table is generated from the same rules the server enforces.

Website changes still need the business’s authorization. An agency’s workspace role does not replace that approval.

Ownership transfers without moving clients, plans or historical records; the previous owner becomes an admin.

Routine reports and alerts follow account preferences and one-click unsubscribe. Security and billing notices are separate.

A weekly email follows a completed check and links its saved answers and coverage.

Monthly reports preserve frozen figures. Added recipients must confirm; an address that refuses cannot be re-added.

Alerts describe supported events with the underlying evidence.

An assistant stopped naming you for a question it used to name you for

Comparable four-week windows: a question previously named you, then none of its latest answers did. Counts and denominators are shown.

The last four weeks are clearly different from the four before

Comparable windows with at least two sampled weeks on each side have non-overlapping approximate confidence intervals. This is not proof of cause or a prediction.

Something new is in the way

A previously passing diagnostic fails, with recorded evidence.

Something that was in the way is fixed

The same diagnostic passes again, with its verification record.

Alert budgets are configurable. Exhausted budgets still record events without sending another alert; comparisons require a consistent method.

Paused or unsubscribed recipients do not receive routine reports. These are service notifications, not newsletters.

New checks and AI work require a paid plan or credit. Vendor requests retain actual or estimated costs.

Plan capacity and optional charges are listed before purchase on the pricing page.