The first source an AI engine lists is not a random draw, and it is not the same thing on every engine. In the Machine Relations Index answer ledger, when the same engine answered the same question twice, the first-listed website of one run was the first-listed website of the other 24.9 percent of the time on ChatGPT, 20.2 on Claude, 17.9 on Perplexity and 11.0 on Gemini. If the order were shuffled each time, those would be 6.0, 7.9, 4.5 and 5.1 percent.
Part of the order is how an engine lays out its list, not which page it likes. ChatGPT put every YouTube video it cited at the end of the list in 615 of the 623 runs that cited YouTube, and none of the 260 lists with five or more sources opened with one. Perplexity lists Reddit, its most-cited website, late: Reddit held the first slot in 3.1 percent of its appearances, where chance is 6.6.
This page reads the order the ledger recorded. It does not show what a person sees on screen, and it does not show that a higher slot earns a click or a citation.
What was counted #
- Population. Successful runs in the Index's buying-question panel, 2026-07-06 to 2026-11-07, five engines (ChatGPT, Claude, Gemini, Google AI Mode and Perplexity) through their APIs, runs that name a brand in the prompt excluded: 197,384 cited-URL events, 23,183 runs and 1,008 questions, the same as the domain-extension study, the URL slug study and the site-sections study. Google AI Overviews and the consumer apps are left out.
- Position. The ledger's recorded citation position, where 1 is the first source in the list the API returned. This page uses only runs that cited five or more sources, so a "first" and a "last" are far apart: 17,115 runs and 179,257 events (90.8 percent of the events). The 6,068 runs with fewer than five sources are excluded; most are ChatGPT (3,842) and Claude (1,642).
- Relative position. (position - 1) / (sources in the run - 1): 0 is first, 1 is last, 0.5 is the middle. A website with no tendency averages 0.5.
- Expected first-slot share. For a website, the average of 1 / (sources in the run) over its appearances: the share of first slots it would hold if order were random.
- Intervals. 95 percent intervals resample the questions with replacement, keeping every run of a question together.
| Engine | Runs with 5+ sources | Events | Questions |
|---|---|---|---|
| Perplexity | 7,765 | 105,251 | 1,008 |
| Claude | 4,337 | 27,298 | 784 |
| ChatGPT | 3,332 | 24,575 | 580 |
| Gemini | 1,359 | 17,287 | 307 |
| Google AI Mode | 322 | 4,846 | 57 |
The engines did not answer the same questions, and Perplexity is 59 percent of the events, so the pooled picture leans toward it. Each engine is read alone below.
The same question gets the same first source more often than chance #
For every pair of runs of the same question on the same engine, the table asks three things. How often is the first-listed website of one run the first-listed website of the other? How often does it appear anywhere in the other run? And, as a baseline, how often does a typical source of one run appear in the other?
| Engine | Questions | First = first | If order were random | First appears in the other run | Typical source appears in the other run | Order correlation |
|---|---|---|---|---|---|---|
| ChatGPT | 580 | 24.9% (23.1 to 27.3) | 6.0% | 39.1% (36.9 to 41.5) | 36.8% (35.2 to 38.6) | 0.57 |
| Claude | 784 | 20.2% (18.3 to 21.7) | 7.9% | 46.9% (44.7 to 48.7) | 35.2% (34.1 to 36.4) | 0.26 |
| Perplexity | 1,008 | 17.9% (16.9 to 19.0) | 4.5% | 47.9% (46.5 to 49.4) | 35.7% (35.0 to 36.5) | 0.39 |
| Gemini | 307 | 11.0% (9.0 to 13.3) | 5.1% | 42.2% (38.1 to 46.4) | 33.2% (30.6 to 35.9) | 0.24 |
| Google AI Mode | 57 | 14.3% (9.6 to 19.2) | 4.5% | 54.4% (49.4 to 59.3) | 33.2% (29.6 to 36.3) | 0.19 |
The order correlation is the correlation of a website's relative position in the two runs, over websites that both runs cited. ChatGPT's order is the most repeatable (0.57) and Gemini's and Google AI Mode's the least (0.24 and 0.19). On Claude, Perplexity and Google AI Mode the first source of a run is also much more likely than a typical source to turn up in a second run (47.9 against 35.7 percent on Perplexity), which fits a first slot that goes to a source the question pulls in repeatedly. Gemini's gap is smaller and its intervals touch (42.2 against 33.2 percent), and ChatGPT's is small (39.1 against 36.8 percent, intervals overlapping), though ChatGPT's first source is the first again a quarter of the time.
Repeated runs of one question were not matched by date, so some of the repeatability can be a source that stayed in the engine's index across the window. Perplexity also changed how many sources it returns in mid-July and on October 7, so the window pools different regimes. Rerun on runs from 2026-10-08 to 2026-11-07 only (79,697 events), the repeat is higher, not lower: the first-listed website was first again in 38.5 percent of ChatGPT pairs, 34.6 on Perplexity, 23.5 on Claude and 15.8 on Gemini, against 7.5, 8.0, 8.4 and 6.7 percent if order were random. Google AI Mode has 63 runs of five or more sources in that window, too few to read. In the same window Reddit on Perplexity averaged 0.65 and held the first slot in 4.7 percent of 127 appearances (chance 10.0), ChatGPT still listed YouTube last (0.99, none first in 131 appearances), Forbes held the first slot in 21.5 percent against 10.0 and The New York Times in 14.3 against 10.2, and PR Newswire on ChatGPT in 18.6 against 12.3, a smaller gap than the full window. No intervals were computed for this check.
Websites that sit early or late on Perplexity #
Perplexity is the engine with enough volume to name websites. Of the 23 websites with at least 250 events in lists of five or more sources, ten have an average relative position whose interval excludes 0.5.
| Website | Events | Average relative position | 95% interval | First slot | Chance |
|---|---|---|---|---|---|
| worldmetrics.org | 281 | 0.34 | 0.30 to 0.38 | 13.2% | 6.8% |
| erpresearch.com | 269 | 0.39 | 0.34 to 0.43 | 7.8% | 7.4% |
| guideflow.com | 337 | 0.42 | 0.38 to 0.46 | 11.0% | 8.2% |
| cnbc.com | 325 | 0.45 | 0.41 to 0.48 | 9.8% | 6.8% |
| nytimes.com | 408 | 0.45 | 0.40 to 0.51 | 16.4% | 7.6% |
| forbes.com | 804 | 0.47 | 0.44 to 0.51 | 12.4% | 6.8% |
| github.com | 375 | 0.54 | 0.51 to 0.57 | 5.3% | 7.7% |
| alibaba.com | 412 | 0.55 | 0.51 to 0.58 | 4.1% | 7.1% |
| youtube.com | 1,348 | 0.55 | 0.53 to 0.57 | 4.8% | 6.8% |
| linkedin.com | 1,108 | 0.55 | 0.53 to 0.57 | 4.4% | 7.1% |
| apple.com | 313 | 0.57 | 0.504 to 0.638 | 4.5% | 7.5% |
| reddit.com | 2,284 | 0.64 | 0.62 to 0.66 | 3.1% | 6.6% |
The New York Times and Forbes are in the table because of the first slot, not the average: the interval for the difference between their first-slot share and chance runs from +2.2 to +15.7 points for the Times and +2.1 to +9.6 for Forbes. Reddit's shortfall in the first slot (3.1 against 6.6) has an interval of -4.4 to -2.6 points. The three user-generated platforms (Reddit, YouTube and LinkedIn) are all listed later than the middle; the business and news publishers are listed earlier. HubSpot, a heavily cited vendor, is at chance (first slot 10.2 against 8.5; interval for the difference -1.4 to +4.9).
ChatGPT lists some websites first and YouTube last #
ChatGPT has fewer websites with enough events, so this table uses 50 or more, and its intervals are wide. These are the clearest departures.
| Website | Events | Average relative position | 95% interval | First slot | Chance |
|---|---|---|---|---|---|
| saasworthy.com | 56 | 0.22 | 0.12 to 0.38 | 60.7% | 12.0% |
| nih.gov | 52 | 0.23 | 0.12 to 0.46 | 51.9% | 16.7% |
| apple.com | 76 | 0.38 | 0.30 to 0.50 | 38.2% | 12.5% |
| businesswire.com | 64 | 0.41 | 0.29 to 0.52 | 31.2% | 12.3% |
| prnewswire.com | 126 | 0.42 | 0.32 to 0.53 | 26.2% | 12.2% |
| techradar.com | 701 | 0.58 | 0.55 to 0.61 | 4.0% | 13.3% |
| axios.com | 55 | 0.76 | 0.67 to 0.85 | 1.8% | 15.4% |
| time.com | 50 | 0.82 | 0.72 to 0.89 | 0.0% | 16.3% |
| youtube.com | 263 | 0.99 | 0.97 to 1.00 | 0.0% | 17.7% |
The intervals for the first-slot gap exclude zero for saasworthy.com (+21.8 to +66.7 points), apple.com (+15.9 to +39.6), PR Newswire (+0.8 to +32.6), Business Wire (+2.7 to +41.4) and nih.gov (+1.9 to +57.8). Each of these websites was cited in lists of five or more sources on only 15 to 21 questions (PR Newswire on 45), so the first slot is a fact about those questions as much as about the website. YouTube is a layout rule: counting every ChatGPT run that cited YouTube whatever its length, 615 of 623 (98.7 percent) had all of its YouTube citations in the final positions. The pattern is ChatGPT's. In lists of five or more sources, 32 of 677 Perplexity runs with YouTube had them all last, and 34 of 329 on Gemini.
On Gemini, Reddit (average 0.59, interval 0.56 to 0.61, 805 events), Medium (0.56) and GitHub (0.59) sit late and Forbes slightly early (0.47, interval 0.41 to 0.52). On Claude, the earliest websites (erpresearch.com, track360.io, impulsec.com) each cited in lists of five or more sources on 16 to 20 questions, hold the first slot in 30 to 34 percent of their appearances against 17 percent by chance. Google AI Mode has too few lists of five or more sources (322 runs) to name websites with confidence.
What this changes #
- Do not read "first citation" as a rank. A first slot means something different on each engine, and part of it is layout. ChatGPT puts YouTube last in nearly every run, and Perplexity tends to list Reddit late.
- Measure position across repeated runs of one question, never one run. The first source repeats only 11 to 25 percent of the time, so a single run's first slot is a coin with a thumb on it. A stable position over many runs is a finding; one run is not.
- A wire release is listed early on ChatGPT, which is not a citation of the company. PR Newswire and Business Wire hold the first slot in 26 and 31 percent of their ChatGPT appearances against 12 percent by chance. That says where ChatGPT puts a wire page when it cites one. It does not say a release makes an engine cite the company that issued it.
- Check whether a visibility report weights slots. A report that counts only the first three sources is counting layout on ChatGPT and Perplexity as well as relevance.
What this page does not show #
- What a reader sees. The order is the one in the API response as the ledger recorded it. The apps may show a different order, hide the list or cap it.
- Any effect of position. The ledger has no click, trust or attention data. A source listed early is not shown to be more persuasive than one listed late.
- Why an engine orders its list as it does. The order may follow retrieval, relevance or the order the answer mentions sources. This page cannot tell which.
- Short lists. Runs with fewer than five sources, most of ChatGPT's and Claude's, are excluded.
- Date effects. Runs of one question were not matched by date in the main tables; the October 8 to November 7 rerun above is a check, not a replacement.
Related parties #
Machine Relations Research is owned by Bramwell, Inc., which also owns AuthorityTech. Both sites are in the data and are counted like any other: in lists of five or more sources authoritytech.io appears 49 times (16 on Perplexity, 21 on Claude, 11 on Gemini, 1 on ChatGPT) and machinerelations.ai 6 times. Neither is near a table threshold, and removing both changes no table on this page.
Method note #
Source: the Index answer ledger's citation-events table, daily export of 2026-11-11 08:10 UTC, read with DuckDB and joined to the run table. Filters: the buying-question panel (mri_v2), runs not flagged as brand-prompted, successful runs, run dates 2026-07-06 to 2026-11-07, engines ChatGPT, Claude, Gemini, Google AI Mode and Perplexity: 197,384 distinct (run, cited URL) events, 23,183 runs, 1,008 questions, 29,282 root domains, reproduced exactly as in the sibling studies before any new count. Position is the ledger's citation_position; no run had two events at one position. Runs with five or more sources only, except the YouTube run count (623 ChatGPT runs of any length that cited YouTube). Websites are root domains. Intervals are percentile intervals from question-level resampling with Python's random module: seed 20261111 and 300 draws for the Perplexity table, 200 draws for the ChatGPT, Claude, Gemini and Google AI Mode scans, seed 99 and 300 draws for the first-slot gap intervals, seed 7 and 200 draws for the repeatability table. The October 8 rerun uses the same definitions on the 2026-10-08 to 2026-11-07 runs, with no resampling. The repeatability table uses every ordered pair of runs of one question on one engine; "first = first" compares the first-listed root domains, and the random baseline is the sum over shared domains of the probability each is first in both runs.
Frequently asked questions #
Which sources do AI engines list first? #
It depends on the engine and the question. On Perplexity, news and business publishers such as The New York Times and Forbes are listed first more often than chance, and Reddit, YouTube and LinkedIn later. On ChatGPT, wire services and a few reference sites are listed first, and YouTube last.
Is the first source in an AI answer random? #
No. When the same engine answered the same question again, the first-listed website was the first-listed website again 24.9 percent of the time on ChatGPT and 17.9 on Perplexity, against 6.0 and 4.5 percent if the order were shuffled each time.
Does ChatGPT list YouTube last? #
Yes. In 615 of the 623 ChatGPT runs that cited YouTube, every YouTube citation sat at the end of the list, and none of the 260 lists of five or more sources began with one.
Does being listed first mean a source is more trusted by the engine? #
The data cannot say. The ledger records the order of the list, not the reason for it, and has no click or trust data. Part of the order is layout, such as ChatGPT's placement of YouTube.
Does the first citation matter for visibility? #
This page does not test it. A visibility report that counts first or top-three sources should say which engine and whether the order was repeated over several runs of the same question, because a single run's first slot repeats only 11 to 25 percent of the time.
Is Reddit listed early or late by AI engines? #
Late. On Perplexity, Reddit's average relative position was 0.64 (0.62 to 0.66) and it held the first slot in 3.1 percent of its 2,284 appearances against 6.6 by chance. On Gemini it averaged 0.59.
See also: how many sources each engine cites, whether engines cite the same sources and Reddit and YouTube.