GEO-114·Issue 10·Content & Semantics·114 / 132
Competitor AI Visibility Benchmarks: Which Rivals to Track
A ten-name tracker mostly records zeros. Track three direct rivals and two substitutes, freeze a 20–30 prompt set, and read three columns every month.
Issue focusMeasurable: AI Visibility, Attribution, and Experiments
In March, a DTC mower brand’s growth lead sits in a Berlin meeting room with a ten-row competitor tracker copied from the Amazon.de shelf. Someone types the question the team sells against: which mower handles 800 square meters on a slope? Three brands come back, two missing from the tracker. The answer ends with “or book a lawn-care service twice a month” — a substitute with no row at all.
A competitor benchmark starts with the roster, not the metrics: three direct rivals decide who replaces you, two substitutes decide whether buyers still buy the category, and a wrong roster measures a battlefield that does not exist.
Why three rivals and two substitutes, not ten
Candidate slots are scarce. One category question names a handful of brands, so a ten-row tracker mostly records zeros, which make a contested category look empty.
Direct rivals answer “replacement” — the names that keep appearing beside yours. Same use case, price band, and channel are filters, not definitions: two near-identical mowers, one sold DTC and one through building-supply chains, never share an answer.
Substitutes answer “displacement”: the cheaper way out before a buyer reaches you — hire someone, rent, buy used, keep the old tool. For mowers that is a per-visit gardener in Germany or a bundled contract in the US. They never enter a comparison page, yet they own the “is this worth buying” layer that sets category size.
| Slot | Who goes there | Test | Failure signal |
|---|---|---|---|
| Direct rival 1–3 | Same use case, price band, channel | Appears beside you most | No shared answer in three months |
| Substitute 1 | Cheaper workaround | “Worth it?” questions | Only in discount questions |
| Substitute 2 | Service or rental | “Don’t want the work” questions | Never appears |
The counterexample is a roster built by fame: the best-known rival may sit two price bands up, named because it owns “premium” while your buyer wants practical.
Lock the prompt set before you sample
Fix the denominator first: four layers — category, scenario, comparison, after-sales — five to eight prompts each, twenty to thirty total, tagged by language, market, and phrasing.
Do not translate a German prompt into English; the corpora differ. Write the buyer’s words: “robotic mower slope” and “will it slide down my hill” return different lists.
Freeze the set for three months. For new prompts, open a new baseline rather than appending rows; a changed denominator breaks the trend. How each layer is scored is covered in the four layers of AI visibility.
Record three columns, not a mention count
Presence: is the brand in the candidate list, 1 or 0. Skip rank in v1; it is the noisiest field.
The quoted fact: the claim of theirs the answer used, usually a line with a unit — “cuts 20 to 60 mm,” “handles 35-degree slopes,” “90 minutes per charge.” That line is the skeleton of their answer unit.
The reason: why the answer recommends them, in three to five words. Slope handling, runtime, cheap, quiet, no upkeep.
A rival gets recommended not for the most specs but because one reason repeats across sources. You cannot copy their specs, but you can check whether you have evidence for it as one quotable sentence.
The mention-count table is the counterexample: the count rises, the candidate list does not. Mentions come from volume; shortlist entry comes from a comparable reason.
The monthly rhythm
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Set the roster | Three rivals, two substitutes, each with “it replaces us when ___” | Content, growth | All five map to a scenario |
| 2 Freeze prompts | 20–30 across four layers, tagged by language and market | Growth | Unchanged three months or re-baselined |
| 3 Sample | Same prompt three times on alternating days; log model, date, region | Growth | Three raw records per prompt |
| 4 Record three columns | Presence, quoted fact, stated reason | Content | Every presence traces to a source sentence |
| 5 Convert to work | Each gap becomes one answer unit or page change | Content | Ships before the next review |
The review takes half an hour: whose presence is rising, at which layer, and which reason we cannot write. Weekly chases noise; quarterly is too slow; one launch reshuffles the list.
Example: one five-row table
Example — brands, specs, and counts are invented. Twenty German prompts, each asked three times on alternating days, sixty samples.
| Slot | Rival | Presence | Quoted fact | Stated reason |
|---|---|---|---|---|
| Direct | Brand A | 31 | Cutting height 20–60 mm, 35° slope | Slope handling |
| Direct | Brand B | 22 | Swappable battery, 90 minutes per charge | Runtime on large lawns |
| Direct | Brand C | 9 | Landed price | Cheap |
| Substitute | Lawn-care service | 17 | Billed per visit, no machine to own | No upkeep |
| Substitute | Corded push mower | 6 | No charging | Tight budget |
Presence first: A and B are the real fight; C appears only on price, which is not the ground to contest.
The quoted facts second: all three carry a unit, none “good quality.” If your pages lack such a sentence, the answer has nothing to carry.
Substitutes third: the service appears 17 times, more than most brands, so nobody owns “no upkeep” — a gap your auto-docking and rain sensing could fill, if written as one sentence.
Next month, ship one or two changes; at five, the sample cannot attribute them.
Where this breaks
First, it is not a sourcing or pricing tool: sixty samples show direction, not share.
Second, the roster expires: new products, channels, and model updates reshuffle answers, so review quarterly and version the baseline rather than trending across a swap.
Third, a brand with no off-site mentions needs its own baseline first; a benchmark shows the gap, not how to fill it, and rivals’ reasons come from reviews and communities mapped in the off-site trust network.
Fourth, rosters differ by market — a per-visit gardener in Germany, a bundled contract in the US — so choose the market with the market priority matrix before naming rivals.
Fifth, in-platform AI cannot be sampled this way: the platform decides the answer, so compare listings, reviews, and fulfillment and keep the two methods apart.
Sixth, never publish the benchmark as a ranking: sampling is not reproducible, and a public list turns observation into a fight.
This month, write five names, each with a sentence: “it replaces us when ___.” Run twenty buyer prompts for two weeks and record three columns. Then take one reason you cannot write, give it a unit, and put it in the answer unit that should own it. Browse the rest on the content and semantics hub.