How Many Sources Compete for AI Citations in Your Category
Measured on 25,323 AI answer runs: a category's citation market holds 362 to 2,525 distinct domains, its top ten hold 8 to 23 percent of citations, and reaching half takes 57 to 250 domains.
Every AI visibility product sells you a leaderboard. Semrush, Ahrefs, Similarweb, BrightEdge, Conductor, Profound, Peec, AthenaHQ and Writesonic all ship a ranked list of domains and a position for yours. The implied instruction is climb into the top ten.
Here is what that top ten is actually worth. Across the 2026-10-14 release of the Machine Relations Index (30,757 cited domains, 25,323 observed answer runs, six engines, 151 days), the ten most-cited domains in a subject category hold between 7.6 and 23.2 percent of that category's citations. Reaching half of a category's citations takes between 57 and 250 distinct domains.
Your category's answer market is not a leaderboard. It is a long tail with a shallow head, and the position worth buying is not the one the dashboards sell.
Key takeaways
- The market is big in every category the Index measures. Distinct cited domains per subject category run from 362 at the small end to 2,525 at the large end, with a median of 1,503.
- The head is shallow. The top ten domains hold 7.6 to 23.2 percent of a category's citations. The most-cited single domain in a category appears in 7.2 to 26.3 percent of its answer runs.
- Entry is cheaper than the dashboards imply. The tenth-ranked domain in a category is cited in 3.0 to 10.1 percent of that category's runs. That is the real price of a top-ten slot.
- Category-local authority is mostly an artifact of the tail. 86.1 percent of all cited domains appear in exactly one category, and 46.0 percent of all cited domains were observed exactly once in 151 days.
- Condition on evidence and it inverts. Among the 381 domains with 50 or more citation observations, the median spans 4 of the 24 categories carrying data, and 17.3 percent are confined to one.
- Both shapes are real. 66 domains cleared 50 observations inside a single category; 70 cleared 50 across ten or more categories.
- Three more categories became measurable. Local Services, iGaming & Betting and Industrial now publish all six question shapes. Local Services is led by two marketplaces, Angi and Thumbtack; Industrial is the flattest market in the Index.
How big is a category's AI citation market?
The Index groups measured questions into subject categories and buyer question shapes. Of its 25 taxonomy nodes, 24 carry citation data in the 2026-10-14 release: 23 subject categories and one legacy bucket of news questions collected before the taxonomy existed. For each, here is how many distinct domains an engine cited at least once, what share the top ten held, how many domains it took to reach half the citations, and the citation rate of the first- and tenth-ranked domain. Ranked rates appear only where the Index publishes a category leaderboard.
| Category | Distinct domains cited | Top-10 share | Domains to reach half | Rank 1 rate | Rank 10 rate |
|---|---|---|---|---|---|
| Legacy News Topics (legacy bucket) | 4,017 | 8.3% | 303 | 10.2% | 4.3% |
| Enterprise Software | 2,525 | 8.9% | 250 | 11.3% | 4.3% |
| AI Visibility & GEO | 2,433 | 13.4% | 153 | 17.0% | 6.4% |
| Cybersecurity | 2,431 | 12.2% | 163 | 13.0% | 5.1% |
| Fintech | 2,382 | 8.5% | 199 | 11.0% | 3.8% |
| Healthcare Services | 2,334 | 9.7% | 227 | 14.9% | 3.4% |
| HR & Talent | 2,073 | 10.4% | 156 | 12.2% | 5.6% |
| Consumer Products | 2,070 | 12.7% | 166 | 16.8% | 4.7% |
| AI Security & Privacy | 1,813 | 13.6% | 125 | 23.0% | 5.8% |
| Martech & Advertising | 1,726 | 12.0% | 137 | 15.7% | 8.3% |
| Consumer Health | 1,674 | 17.6% | 98 | 26.3% | 7.3% |
| Deep Tech & Hardware | 1,505 | 13.3% | 129 | 13.0% | 5.8% |
| AI Infrastructure | 1,503 | 18.4% | 89 | 17.5% | 6.5% |
| Education & Training | 1,476 | 12.5% | 135 | 14.0% | 4.7% |
| Consumer Finance | 1,326 | 23.2% | 57 | 24.9% | 7.5% |
| Emergent Prosumer | 1,279 | 13.9% | 140 | 17.6% | 4.5% |
| Industrial | 1,240 | 7.6% | 205 | 7.2% | 3.0% |
| iGaming & Betting | 1,075 | 12.3% | 95 | 15.9% | 4.8% |
| Local Services | 1,044 | 16.4% | 113 | 16.8% | 5.7% |
| Family Software | 941 | 22.8% | 58 | 17.0% | 10.1% |
| Logistics & Freight | 833 | 12.4% | 108 | 13.0% | 7.3% |
| Professional Services | 471 | 10.1% | 122 | collecting | collecting |
| Physical Consumer Products | 454 | 14.7% | 104 | collecting | collecting |
| Sales, GTM & CRM | 362 | 15.3% | 73 | collecting | collecting |
Read the third column first. In Consumer Finance, the most concentrated category, 57 domains carry half the citations. In Enterprise Software it takes 250. A leaderboard that shows you ten rows is showing you well under a twentieth of the domains that build half the answers in Enterprise Software, and about a sixth of them in Consumer Finance. Those are different businesses, and the same dashboard renders them identically.
What does it actually cost to reach the top ten?
Less than the framing implies, and it buys less than the framing implies.
The tenth-ranked domain in Industrial was cited in 3.0 percent of that category's runs. In Family Software, the most concentrated small category, it was 10.1 percent, the only category where the tenth slot clears one run in ten.
Here is Cybersecurity's actual head, with each domain's share of the category's 1,771 observed runs:
| Rank | Domain | Runs cited | Category citation rate |
|---|---|---|---|
| 1 | sentinelone.com | 231 | 13.0% |
| 2 | paloaltonetworks.com | 230 | 13.0% |
| 3 | reddit.com | 182 | 10.3% |
| 4 | microsoft.com | 178 | 10.1% |
| 5 | huntress.com | 149 | 8.4% |
| 6 | crowdstrike.com | 119 | 6.7% |
| 7 | cynet.com | 114 | 6.4% |
| 8 | gartner.com | 108 | 6.1% |
| 9 | expertinsights.com | 104 | 5.9% |
| 10 | linkedin.com | 90 | 5.1% |
Six of the ten are security vendors cited through their own domains. Two are general platforms, one is an analyst firm and one is a market database. The category leader appears in roughly one run in eight. Holding rank 10 in Cybersecurity means being cited in about 90 of 1,771 observed runs: a real target with a number attached, which is the opposite of what a percentile score gives you.
Fintech and HR & Talent split the other way from each other. Fintech's head is vendors: stripe.com at 11.0 percent, linkedin.com at 8.9, airwallex.com at 8.2, chargeflow.io at 4.8, adyen.com at 4.7. HR & Talent is led by an editorial review publication: peoplemanagingpeople.com at 12.2 percent, then rippling.com at 8.3, forbes.com at 7.5, g2.com at 6.8 and linkedin.com at 6.7. Which kind of source holds the head decides whether earned coverage or your own pages is the faster route in.
What do the newly measurable categories show?
Local Services, iGaming & Betting and Industrial carried citations in September but no question shape above the evidence floor. All three now publish every shape, and they could hardly look less alike.
- Local Services is a marketplace market. angi.com is cited in 16.8 percent of the category's 631 runs and thumbtack.com in 15.8 percent, ahead of reddit.com at 11.1. Directories and review sites (expertise.com, consumeraffairs.com, yelp.com, bbb.org, homeadvisor.com) fill most of the rest of the top ten. A local provider is described to the engine by the platforms that list it.
- iGaming & Betting is a supplier market. The head is B2B platform and software suppliers: track360.io at 15.9 percent of 800 runs, gammastack.com at 11.1, smartico.ai at 9.6. actionnetwork.com at 5.2 percent is the only betting media domain in the top ten.
- Industrial is the flattest market in the Index. Its leader, linkedin.com, appears in 7.2 percent of 626 runs, its top ten hold 7.6 percent of citations, and half the citations take 205 domains, more than any subject category but Enterprise Software and Healthcare Services.
Is AI citation authority category-specific?
This is where the data sets a trap.
The headline number says yes: 86.1 percent of the 30,757 cited domains appear in exactly one category. That reads as a clean finding: authority is local, win your niche, breadth is a myth.
It is an artifact of the tail. 46.0 percent of all cited domains were observed exactly once across 151 days. A domain cited once can only ever appear in one category. The 86.1 percent is mostly counting domains that brushed an answer a single time, and it says nothing about the sources engines rely on.
Condition on evidence and the relationship reverses:
| Minimum citation observations | Domains | Single-category share | Mean categories |
|---|---|---|---|
| 1 or more | 30,757 | 86.1% | 1.27 |
| 2 or more | 16,622 | 74.2% | 1.50 |
| 5 or more | 7,036 | 60.7% | 1.93 |
| 10 or more | 3,323 | 49.8% | 2.47 |
| 25 or more | 1,044 | 33.6% | 3.72 |
| 50 or more | 381 | 17.3% | 5.65 |
The 381 domains with 50 or more observations hold 28.7 percent of all 167,832 citation observations in the release. Their median spans 4 of the 24 categories with data. The Index grades domains A, B or C by how much evidence stands behind their rate; 20 domains currently hold grade A, and every one of them appears in more than one category.
Grade C, a much larger set at 679 domains, is 32.1 percent single-category. So the pattern is monotone: the more an engine leans on a source, the more categories that source turns up in.
Two readings survive this, and they are both true. Domain authority in the classic link-graph sense still buys you nothing with an answer engine; that case rests on a different measurement, and Machine Relations Research has published the inversion. And breadth of observed citation travels with depth of it. Those are not in conflict. The first is about what earns a citation. The second is about what having earned many of them looks like from the outside.
Can a single-category source still win?
Yes, and 66 of them did. These are the ten with the most citation observations among domains that cleared 50 while appearing in exactly one category:
| Domain | Category | Citation observations | Confidence grade |
|---|---|---|---|
| allaboutcookies.org | Family Software | 138 | B |
| omnimd.com | Healthcare Services | 128 | B |
| useboomerang.com | Family Software | 118 | B |
| cynet.com | Cybersecurity | 114 | B |
| impulsec.com | Family Software | 110 | B |
| angi.com | Local Services | 106 | B |
| anysilicon.com | Deep Tech & Hardware | 104 | C |
| ubiehealth.com | Consumer Health | 104 | C |
| safewise.com | Family Software | 102 | B |
| thumbtack.com | Local Services | 100 | B |
Against those, 70 domains cleared 50 observations across ten or more categories. The widest are the ones you would guess: reddit.com in 24 categories on 2,429 observations, youtube.com in 24 on 1,812, linkedin.com in 23 on 1,453, forbes.com in 21 on 1,017, medium.com in 19 on 951, nih.gov in 17 on 629.
The practical read: a deep single-category position is achievable and it tops out around 140 observations, while the broad positions run into the thousands. If your category is the whole business, the specialist path is the one with a realistic finish line. If you are trying to be a source engines reach for by default, breadth is what that looks like when it works.
Which categories are not measurable yet?
The Index publishes a citation rate for a category and question shape only after that segment clears an evidence floor of 10 observations across 7 distinct run dates. Below the line it is marked collecting rather than scored, which is a statement about the evidence and not about the market.
Three of the 23 subject categories (Professional Services, Physical Consumer Products, and Sales, GTM & CRM) have citations but no question shape above the floor. Martech & Advertising has one of seven shapes scored and Logistics & Freight two of six. If someone sells you a category score in one of those five, ask what evidence stands behind it, because the Index's evidence is still accumulating. VC & Private Equity carries no citation data in this release at all.
What to do with this
1. Replace the percentile with a run count. Rank 10 in your category is a number of observed runs, not a grade. In Cybersecurity it is 90. Ask your vendor for the denominator behind any score they show you; a rate without an observation count is a shape, not a measurement. Sielinski's March 2026 uncertainty work is the reason this matters: rankings move between repeated samples, so a position that is not carrying its evidence count is not carrying anything.
2. Aim at the half-line, not the head. The domains that reach half your category's citations number between 57 and 250. That set is where the answer actually gets built. Being inside it is a reachable goal in a way that displacing the category leader usually is not.
3. Know which source class holds your head. Six of Cybersecurity's top ten are vendors cited through their own domains; Local Services is held by marketplaces; HR & Talent by an editorial review publication. Engines pull from whatever answers the question directly, and the mechanics of how each engine selects and grounds sources are documented by OpenAI, Perplexity, Anthropic, Google and Microsoft, and each behaves differently.
4. Do not spend on breadth you have not earned. Breadth is a consequence of volume in this data, not a substitute for it. Domains appearing in many categories got there by being cited a lot somewhere first.
5. Measure on the surface that owns the answer. Per-engine behaviour diverges enough that a single blended number hides the thing you need. Read each engine on its own. Google's own documentation on AI features in Search and its AI Mode announcement describe surfaces with different source behaviour again.
Frequently asked questions
How many domains compete for AI citations in a typical category?
The median subject category in the 2026-10-14 Machine Relations Index release had 1,503 distinct domains cited at least once across 151 days of observation. The range across the 23 subject categories carrying data was 362 to 2,525.
What share of AI citations do the top ten domains hold?
Between 7.6 and 23.2 percent, depending on the category. Consumer Finance was the most concentrated at 23.2 percent; Industrial the least at 7.6 percent.
Is a top-ten AI visibility ranking worth chasing?
It is worth translating. A top-ten slot means a citation rate between 3.0 and 10.1 percent of a category's observed runs. That is a concrete target you can price. The rank by itself carries no information about how far away it is.
Do AI engines cite the same sources across different categories?
For the sources they lean on, yes. Among domains with 50 or more citation observations the median spans 4 of 24 categories. Across the full index 86.1 percent appear in a single category, but 46.0 percent of all cited domains were observed only once, which is what drives that figure.
Which sources do AI engines cite for local services questions?
Marketplaces and directories. In the 2026-10-14 release angi.com is cited in 16.8 percent of 631 Local Services runs and thumbtack.com in 15.8 percent, ahead of reddit.com, expertise.com, consumeraffairs.com, yelp.com and bbb.org.
Why do some categories have no published citation rate?
The Index requires 10 observations across 7 distinct run dates before it publishes a rate for a category and question shape. Professional Services, Physical Consumer Products and Sales, GTM & CRM are still accumulating evidence, so their segments are marked collecting rather than scored.
Sources and method
Every figure above is computed from the public Machine Relations Index release of October 14, 2026: 30,757 cited domains, 25,323 observed answer runs, six engines, observation window 2026-05-10 through 2026-10-14, evidence floor of 10 observations across 7 run dates. The public dataset reports citation rates, rankings and evidence counts; it excludes raw cited URLs and provider payloads, and its methodology and update log is public.
A citation observation is one domain cited in one observed run of one category and question shape, summed from each domain's published segment counts. Per-category denominators are the sum of observed runs across that category's question shapes, and across all 24 nodes they total exactly the release's own published run count of 25,323. Category counts are of domains with at least one citation observation in that category. "Domains to reach half" is the smallest number of domains whose citation observations sum to at least half the category total. Rank 1 and rank 10 rates are read from the Index's published category leaderboards, never re-derived. Headline ranges cover the 23 subject categories and leave out the legacy news bucket. Confidence grades A, B and C are the Index's own evidence tiers.
This page was first published on 2026-09-23 from that day's release, which carried 16,475 runs and a different methodology version (methodology v2.0). Every figure was recomputed on 2026-10-14, so the two versions are not a time series; part of any difference comes from the method. The single-category share is driven by the once-observed tail, which is why the conditioned table is the one that carries the argument.