How we measure Active Competitors
A scan asks every selected engine your question bank, reads the brands out of each answer, merges spellings of one company, and reports how much of the conversation each brand holds. It is the only audit allowed to discover competitors rather than measure ones you named.
Depth: wordings per question
Draw one is your question as written; the rest are meaning-preserving rewordings in your market's language, asked on every engine you selected and pooled under the question.
| Wordings per question | Grade | Cost | What it buys |
|---|---|---|---|
| 1 Default | T1SnapshotOne read. Rerun to confirm it. | ×1 | |
| 2 | T2Directional*Shows the direction. The rate can move. | ×1.9 | Of 207 questions, a quarter split at two wordings. |
| 3 | T3Consensus*Rerun, and most reads matched. | ×2.8 | Of 174 question groups, none split. |
| 5 | T3Consensus*A rate within 30 points either way. | ×4.6 | Plus or minus 25 points. |
| - | T4Established | - | Not offered. |
| - | T5Verified | - | Not offered. |
Why it stops at T3. Share of voice stays near plus or minus 25 points at every depth.
* Our estimate, not yet validated by real runs.
The first rung is unstarred because it is a property of the code, not an estimate: one extraction read and one merge pass produced every brand, so a single-wording scan reads Snapshot even over two hundred answers. There is no Established rung because answers to one question are not independent. Their correlation is 0.187 across 2,270 answers, so at a typical instrument the effective sample size runs 11.96 at one wording and 12.73 at five.
What depth buys here is discovery. Template 28, run three times inside eleven minutes on 18 September 2026, found 12, 34 and 31 brands from the identical question set. Only engine fetches multiply with depth: the rival teardown is a flat fee run once on the final brand set, and the one Google result set per question is counted off your base bank. A wording we cannot make genuinely different is not run and not charged, and the report states how many wordings were asked.
Breadth: questions per template
How many questions the template asks. This is the one ladder in the product with no Saturated rung, because the data says so.
| Questions | Grade | Cost | What it buys |
|---|---|---|---|
| - | T1 | - | Not on the scale. Narrow is the lowest breadth grade. |
| 1 | T2NarrowNew finds had not slowed yet. | ×1 | |
| 3 | T2NarrowNew finds had not slowed yet. | ×3 | |
| 8 Default | T2Narrow*New finds had not slowed yet. | ×8 | Median of 86 rivals found. |
| 20 | T3Partial*Still finding new things at the end. | ×20 | |
| - | T4Mostly saturated | - | Not offered. |
| - | T5Saturated | - | Not offered. |
Why it stops at T3. Coverage stayed near 44% of rivals at every size measured.
* Our estimate, not yet validated by real runs.
The curve was rebuilt over all 44 stored scans. From 15 answers to 229, sample size correlates negatively with estimated coverage and strongly positively with rivals found: a wider scan finds more competitors and a smaller share of them. So the percentage never appears without its frame.
Every scan stamps the grade its own curve earned, often better than the rung promised. The curve's grade and the coverage percentage travel together and where they differ the lower one governs. The percentage is a lower bound on how many competitors exist, so real coverage is at best what we show; and it counts distinct names, so two spellings the merge failed to join count as two.
Seven sources of uncertainty
Share of voice is the end of a seven-stage chain. Our interval used to price one stage and read as covering the run: a published 95% interval contained a re-run of the same template at the same size only 44 times in 64, and template 28's three runs returned 77.8%, 72.2% and 86.7%.
| Source | Priced | What could have gone differently |
|---|---|---|
| 1. Which questions were asked | Always | Your bank is a sample of what buyers ask. The cluster bootstrap redraws your intent groups 1,000 times, labelled as covering this stage only. |
| 2. How each question was worded | At 2+ wordings | Rewording moves the answer more than rerunning does. |
| 3. What the engine answered this time | At 2+ wordings | The same question asked twice returns different prose, and a brand can drop out. |
| 4. Which answers arrived at all | Never | Failed calls leave the denominator. The report states completed out of attempted. |
| 5. Which brands the extractor named | At 2+ wordings | A failed batch removes brands from answers that stay in the denominator. |
| 6. Which surface forms were merged | At 2+ wordings | An over-merge inflates a share and a split halves it, and both look identical on the report. |
| 7. Whether we recognised you | Never | Matched by name keys and domain. A miss reports 0%, so a run with no match shows no interval at all. |
Sources 2, 3, 5 and 6 sit inside one draw, so a second wording re-rolls all four. The no-match rule replaced a plus or minus 0.00: nine of 44 stored scans reported a certain zero that way, and at least one was contradicted by the next run of the same template two days later. The market check is absent on purpose: it removes out-of-market rivals and changes your rank, but share of voice divides answers naming you by answers completed, so it cannot move the number.
What the report claims
| Number | Type | What it can support |
|---|---|---|
| Brand extraction | Interpreted | A model reads each answer and names the companies. A judgment, not a string match. |
| Whether an answer named a known brand | Counted | A fact about a list we already hold. |
| Co-mention | Counted | Two brands in one answer are associated, not endorsed. We count it and claim nothing about what it means. |
| Market verdict, written summary | Interpreted | Labelled as model judgments. |
| Share of voice | Derived | Never more certain than the weakest thing it was computed from. |
Fidelity: what is read directly
The answers are real: every one is a live engine call, stored verbatim and readable on the report. The brand list inside them is one remove away, because a model reads each answer and reports the companies it found. That trade is deliberate: a string matcher cannot recognise a company it was never told about, and discovery is the point.
Two things come from outside the answers, both read from the real surface and neither in the share-of-voice path: each discovered brand's homepage, fetched to judge whether it operates in your market, and one Google result set per question for the organic-ranking column. An earlier version counted capitalised words in a field called competitor_mentions, which is how "The" reached a published report; that field is gone.
What we do not claim
- That we found every competitor. On the evidence above we usually found under half the names.
- That the engines are right about who competes with you.
- A confidence percentage on a model judgment. Extraction and the market verdict are graded on agreement between runs, and agreement is not accuracy: a model confidently wrong five times out of five agrees perfectly.
- That the interval covers the run. It covers the aggregation stage and says so beside itself.
- Anything but one wording for scans run before sampling existed. That is what they were, and we have not re-graded them.
