Every AI Engine Names Different Competitors for Your Brand
Build a competitor list from one AI engine and roughly three quarters of it is that engine's opinion alone. Across 2,317 company names returned for 26 brands, only 12.5% were named by all three engines. The ordering is worse: in most cells the leading rival is ahead by a single mention. Treat the set as a set.
Ask an AI engine who competes with you and it will hand back a list. Ask a second engine and you get a different list. Not a reordering of the same names, a different list.
Across 2,317 company names returned as competitors for 26 brands, 73.8% were named by exactly one engine. Only 12.5% were named by all three.
So a competitor list assembled from a single engine is roughly three quarters that engine's opinion. The practical fix is cheap and almost nobody applies it: treat agreement across engines as the filter that separates a competitive set from a list of names.
What We Measured
Every company named as a competitor in a tracked answer between 2026-07-06 and 2026-08-22, for the brands that ran on all three engines in that window. Names are normalised and counted once per brand per engine, so a company mentioned in six answers on ChatGPT counts as one ChatGPT name, not six.
One methodological choice does the heavy lifting. We count distinct names across answers, never names per answer. Our parser caps how many companies it records from a single response, so a per-answer count would describe that ceiling rather than the engine. Counting whether a name appeared at all is immune to it: a company recorded eighth in one answer still counts once.
The window starts in July because our engines moved to the current generation of web-search models around then, and the corpus before that cutover describes retired ones. The same cutoff governs our per-engine citation work.
The Corroboration Split
How many engines named each company as a competitor
2,317 distinct company names returned across 26 brands, 2026-07-06 to 2026-08-22. Counted once per name per brand, so a name mentioned repeatedly in one answer is not counted twice.
- All three
- 12.5% · 290 names
- Two
- 13.7% · 317 names
- One only
- 73.8% · 1,710 names
Counting distinct names across answers rather than names per answer is deliberate: our parser caps a single answer's list, so a per-answer count would measure that ceiling instead of the engines.
The long grey bar is the finding. 1,710 of the names, 73.8%, came from a single engine and were never echoed by another. The corroborated core is small: the median brand had just 12.3% of its named competitors confirmed by all three.
Some of that tail is real. Engines genuinely weight different sources, so one may surface a regional rival another has never encountered. Some of it is not real at all: a plausible-sounding company name produced once and never again. The useful part is that you do not have to tell those two apart by hand. A second engine does it for you, and a third settles the argument.
Why the Engines Disagree
The mechanism is not mysterious, and we measured it separately. Engines draw on citation pools of very different widths: half of ChatGPT's citations come from just 147 domains, against 521 for Gemini, with Perplexity in between. That is the same measurement that explains why Reddit dominates one engine and barely registers on another.
A competitor list is downstream of that. If two engines are reading substantially different sets of pages about your category, they will name substantially different companies, and neither is malfunctioning. A narrow pool keeps returning the same well-covered names. A wider one reaches further and turns up companies whose coverage is thinner.
That predicts the shape we found. The corroborated core is the set of companies covered well enough to appear in everyone's sources. The single-engine tail is companies present in one pool and absent from the others, plus a residue of names that were never real to begin with.
It also means the disagreement is not a defect to be averaged away. Each engine is telling you something true about the slice of the web it reads. The question worth asking is which slice your buyers are getting when they ask.
You Cannot Rank Them
The obvious next move is to rank the set: who is my biggest rival on each engine. We tried it and the answer does not survive contact with the data.
How far ahead the most-named rival actually is
Margin between the most-named and second-most-named competitor, across 96 brand-engine cells in the same window. Measured in mentions, not percentages.
- Dead tie
- 26%
- Ahead by 1
- 32%
- Ahead by 2
- 18%
- Ahead by 3+
- 24%
The "ahead by 2" band is the residual once the other three are counted, so it carries their rounding. For contrast, the top-three rival SETS overlap between engine pairs by a median of 50%. The membership carries information; the ordering inside it mostly does not.
Across 96 brand-engine cells, the median gap between the most-named competitor and the runner-up is 1 mention. 26% are outright ties, and 58% sit within a single mention. Only 24% have a leader clear by three or more.
A one-mention margin is not a ranking, it is a coin flip with extra steps. That fits what we measured separately about run-to-run movement: if a whole visibility score can move fifteen points between runs with nothing changed, a single mention certainly can. Any leaderboard built on these margins will reshuffle next week and invite someone to explain why.
We could have led this post with the version that tests well: only six of the twenty-six brands had the same top rival on every engine. It is arithmetically true and it means almost nothing, because the leaders it compares were mostly decided by one mention. Reporting the instability is the more useful finding.
What Does Hold Up
Membership, at the set level. Comparing each brand's top three rivals between pairs of engines, the median overlap is 50%. Two brands of 26 shared nothing at all across engines and four shared everything.
Half is a genuinely informative number. It is far too high to call the engines random: they are clearly describing the same market. It is far too low to treat any one of them as the market. If your competitive picture comes from whichever engine you happened to check, you are working from half a map and cannot tell which half.
This matters most for a metric people quote without thinking: share of voice is a ratio whose denominator is the competitor set. Let the long tail of single-engine names into that denominator and your share falls for reasons that have nothing to do with your visibility.
Squaring With the Count Post
We published a related finding earlier: the number of competitors AI names never stops growing. Sample the same brands harder and the count climbs, because the count was describing how often we asked rather than the market.
That result predicts this one. If deeper sampling keeps turning up new names, most of those names should be thinly attested, and they are: 73.8% appear on one engine only. The two posts are the same phenomenon seen from different sides. The count is unstable because the membership is mostly tail, and the tail is mostly single-engine.
Which is why neither post ends at "the number is wrong". Corroboration turns an unbounded list into a bounded one. Our own product already drops names seen once with no resolvable domain inside a single window; agreement across engines is a stronger version of the same test, and it is one you can run by hand.
What to Do With This
- 01Never build the list from one engine. Roughly three quarters of what you get back will be that engine alone, and nothing on the page tells you which names those are.
- 02Use agreement as the filter. Names confirmed by two engines are worth a look; names confirmed by three are your real competitive set. Expect that set to be small.
- 03Do not rank them. Report the set and each member's presence over time. A leaderboard built on one-mention margins will reorder itself weekly.
- 04Audit your share-of-voice denominator. If it includes every name any engine ever produced, the number is measuring the tail rather than your standing.
- 05Then run the audit on the survivors. The seven-step competitor audit is worth its effort against a corroborated set and wasted against a list of one-offs.
One limit worth stating. This is 26 brands over seven weeks, and the engines each brand runs are set by its plan, so the cohort is not a random sample of the market. The direction is consistent across every brand in it, but treat the exact percentages as this cohort's, not the industry's.
See which rivals every engine agrees on
Ranqo records every competitor named in every tracked answer, per engine, so the corroborated set separates itself from the long tail without you reading transcripts. The free checker runs unbranded prompts from your category against ChatGPT live and shows who else comes up. No signup.
Run the free checkWritten by
Nisha Kumari
Nisha Kumari is Co-Founder at Ranqo, where she leads growth strategy and client acquisition. With a background in digital marketing and financial management, she specializes in SEO, Generative Engine Optimization, and helping brands build visibility across AI platforms.
Share this article
Related articles
Your AI Competitor List Never Stops Growing
We nearly published a finding: AI names about 91 companies as the average brand's competitors, and only 18 recur. Then we sampled the same brands harder and 91 became 282, still climbing. The count was describing how often we asked, not the market. Here is the test that settles it, and what it means for every share-of-voice number.
How to Audit a Competitor's AI Visibility (and Find the Pages Doing the Work)
Most competitor audits stop at a spreadsheet of who got mentioned, which tells you a rival is winning without telling you why. Only 2.9% of AI citations point at a brand's own domain, so the page carrying your competitor almost always belongs to someone else. Here is the seven-step audit that finds it, and the part no audit can fix.
Your AI Visibility Score Moves 15 Points for No Reason
We measured what an AI visibility score does when nothing happens. Between consecutive runs, with no fix shipped, the overall score moves up to 14.5 points and a single category 27.3. The narrower the slice, the noisier it gets, which is the opposite of what everyone assumes. Most reported wins in this category are sampling variance.