Asking the same question again already changes the sources an AI engine cites. Rewording the question changes them a little more, and broadening or narrowing it changes them far more. In the Machine Relations Index question panel (980 questions and 22,620 answer runs, July 6 to October 20, 2026), each question's runs were split at random into two halves. One half shared 46.1 percent of its cited-source volume with the other half of the same question, 39.4 percent with a half of a reworded version of the question, and 23.0 percent with a half of a broader or narrower version at the same intent.
A run is one question asked of one engine on one day. Shared volume is the overlap between two sets of runs: for every cited domain, take the smaller of its share of each set's citations, then add those up. Identical citation mixes score 100 percent, and sets with no domain in common score 0. Questions with a different intent stage, a different question shape or a different category share 7.8, 6.1 and 1.5 percent (rows defined in the table below). This page reads that ladder for one decision: how much a single tracked prompt, and a single wording of it, can tell you when measuring AI visibility.
What was compared #
- Panel. The Index's 980 buyer questions in 24 categories and six question shapes (best tools, how to choose, is it worth it, problem-first research, top lists and head-to-head comparisons), with news-driven questions left out. They ran on six engines across 103 run dates, 22,620 runs in all, an average of 23.1 per question. Of those runs, 16,975 cited at least one source, adding up to 135,864 domain-run pairs. A domain counts once per run.
- Halves. The runs of each question were split into two random halves, by the parity of a hash of the run identifier, so both halves hold the same mix of engines and dates in expectation. A half holds about 12 runs on average. Every comparison sets one half against one half, and a half needs at least 8 domain-run pairs to count. 894 of the 980 questions qualified for the rerun comparison.
- Reworded. The Index builds 280 pairs of questions that are worded differently and share a category, a question shape and an intent stage. Each comparison sets a half of one question against the opposite half of its partner.
- Broader or narrower. 140 intents are each asked at three levels of specificity: a middle wording, a broader one and a narrower one that adds constraints. The 790 comparisons mix the middle wording against its broader version, the middle against its narrower version, and the broader against the narrower, a third each by design.
- Further apart. A different intent stage in the same category and question shape, a different question shape in the same category, and a different category with the same question shape and intent stage (three sampled partners per question).
The ladder: how much cited volume survives each change #
| Compared with the question's own half | Comparisons | Shared citation volume | Shared domains (Jaccard) | Citations per half, average |
|---|---|---|---|---|
| Other half of the same question (a rerun) | 894 | 46.1% | 0.335 | 75.0 |
| Half of a reworded question | 511 | 39.4% | 0.276 | 64.9 |
| Half of a broader or narrower question, same intent | 790 | 23.0% | 0.147 | 86.4 |
| Different intent stage, same category and shape | 4,135 | 7.8% | 0.049 | 73.1 |
| Different question shape, same category | 31,492 | 6.1% | 0.038 | 74.7 |
| Different category, same shape and intent stage | 5,368 | 1.5% | 0.009 | 72.9 |
Jaccard counts each domain once: it is the share of the distinct domains cited in either half that appear in both.
The rerun row is the reference point for this measure. Two halves of the same question, asked of the same engines over the same weeks, share under half of their cited volume, and about one in three of the distinct domains they cite between them appears in both. Every other row starts from that level, not from 100 percent.
Rewording costs about 7 points on top of the rerun. Broadening or narrowing costs about 23. A different intent stage in the same category falls to 7.8 percent, so a broader or narrower version of a question still behaves much more like the original than a different question in the same segment does.
Reworded and broader or narrower, pair by pair #
The ladder rows cover different questions and different half sizes. To compare like with like, the table below keeps only the pairs where both questions have a rerun score of their own, then sets each pair's overlap against the rerun level of those same two questions.
| Pair type | Matched pairs | Rerun level of the same questions | Level across the two versions | Share of the rerun level kept |
|---|---|---|---|---|
| Reworded | 244 | 45.3% | 39.8% | 88.0% (95% interval 84.7 to 91.5) |
| Broader or narrower | 374 | 49.0% | 23.5% | 47.9% (95% interval 44.4 to 51.9) |
The intervals come from resampling the 24 categories 1,000 times. A separate program that recomputed the ladder from the run records with five equal-size random splits gave 87.4 to 88.2 percent for reworded and 47.2 to 48.2 percent for broader or narrower.
The split rule moves the level, not the order #
How a question's runs are divided into halves changes the rerun level, so the page reports three rules.
| Split rule | Rerun | Reworded | Broader or narrower | Reworded keeps | Broader or narrower keeps |
|---|---|---|---|---|---|
| Random halves (headline) | 46.1% | 39.4% | 23.0% | 88.0% | 47.9% |
| Alternating run dates | 53.5% | 39.8% | 24.2% | 77.0% | 43.5% |
| Earlier runs against later runs | 42.3% | 33.5% | 21.9% | 83.8% | 49.1% |
Other random splits stay close to the headline. Across the headline split, two other hash salts and an engine-stratified split, the rerun level is 46.1 to 46.4 percent and the reworded level 39.1 to 39.6 percent. The order, rerun above reworded above broader or narrower, holds under every rule. The honest summary is a range: a reworded question keeps 77 to 88 percent of the rerun overlap, and a broader or narrower one keeps 44 to 49 percent. This page does not establish why the alternating-date rule gives the highest rerun level.
Is it a few domains, one engine or one category? #
- Without the head. Removing the 25 domains ranked highest in the release leaves 45.3, 38.2 and 21.3 percent for rerun, reworded and broader or narrower.
- Three engines. Using only ChatGPT, Claude and Perplexity gives 46.3, 38.9 and 22.0 percent.
- Categories. Rerun above reworded above broader or narrower holds in 21 of the 24 categories. The three exceptions (professional services, sales and GTM CRM, and venture capital and private equity) have 12, 12 and 5 rerun comparisons.
- Single engines. The order holds inside each engine that has enough runs for the ladder. Halves are built from that engine's own runs and are smaller than the pooled halves, so these rows show that the order holds, not which engine is steadier or how it compares with the pooled table.
| Engine | Rerun | Reworded | Broader or narrower | Domains per cited run | Citations per half, average |
|---|---|---|---|---|---|
| Perplexity | 47.0% | 37.8% | 22.0% | 12.7 | 40.3 |
| ChatGPT | 48.9% | 32.0% | 12.2% | 4.7 | 16.5 |
| Claude | 43.6% | 37.6% | 18.7% | 5.1 | 15.7 |
Comparison counts per engine, in the order rerun, reworded, broader or narrower: Perplexity 806, 470 and 736; ChatGPT 579, 328 and 531; Claude 572, 315 and 509. Gemini is counted in the pooled table only: it returned a citation in 1,262 of its 5,023 runs in this panel, and the release manifest lists it with status degraded. Google AI Mode and Google AI Overviews ran on about 140 questions each (817 and 644 runs), too few for a ladder of their own. The pooled table is weighted by citation volume, and Perplexity supplies 50.7 percent of the 135,864 domain-run pairs, against 17.2 for ChatGPT, 15.8 for Claude, 10.0 for Gemini, 3.4 for Google AI Mode and 2.9 for Google AI Overviews, which is consistent with the Perplexity row sitting close to the pooled one. For why engines are observed on different numbers of days, see per-engine observation coverage.
One run against a batch of runs #
The ladder compares halves of about a dozen runs. A single run is a much smaller sample: a run that cited anything cited 8.0 domains on average (135,864 pairs over 16,975 cited runs). Across 188,338 pairs of cited runs of the same question, two single runs shared 14.3 percent of their cited volume (Jaccard 0.111). Pairs from the same engine, necessarily on different days, shared 31.6 percent (43,839 pairs, Jaccard 0.248) and pairs from different engines 9.0 percent (144,499 pairs, Jaccard 0.069).
Rerun overlap rises with the size of the half. Grouped by the number of citations in the smaller half of each rerun comparison, it is 35.5 percent for 8 to 39 citations (302 comparisons), 50.6 percent for 40 to 79 (391), 52.2 percent for 80 to 159 (170) and 59.3 percent for 160 or more (31).
A domain's appearance in a few runs is also weak evidence. Among the 658 questions with at least 8 runs in each half (14.5 runs per half on average), the table shows how often a domain cited in the first half was cited again in the other half.
| Runs citing the domain in the first half | Question-domain pairs | Cited again in the other half |
|---|---|---|
| 1 | 19,828 | 37.8% |
| 2 | 6,098 | 64.7% |
| 3 to 4 | 3,965 | 83.2% |
| 5 or more | 1,813 | 95.6% |
| All | 31,704 | 52.0% |
Recomputing this with other random splits moves each row by up to two points.
What this means for choosing and tracking prompts #
- Count runs before counting wordings. A rerun alone leaves more than half of the cited volume different, and two single runs share 14.3 percent. A tracked prompt supports a reading about its sources only when the run count behind it is disclosed (see run counts behind engine readings).
- Wording is a second-order change. A reworded question keeps 77 to 88 percent of the rerun overlap, so tracking one intent through two or more wordings adds breadth while behaving much like extra runs. A single wording is one draw of the wording as well as one draw of the runs.
- A broader or narrower question is another question. Keeping 44 to 49 percent of the rerun overlap puts it far below a reworded one and far above a different intent. Track broader and narrower versions of an intent as separate questions with their own run counts, and do not pool them as if they were reruns.
- Read one appearance cautiously. A domain cited in one run of a half was cited in the other half only 37.8 percent of the time, so a single appearance is weak evidence either way.
- Report the rule. A reading should state its question set, its run count and, where it compares periods, how runs were assigned to periods, since the same data gives a rerun level from 42.3 to 53.5 percent depending on the split.
What this page does not show #
- Customer prompts. The panel is the Index's own question set, written to cover buyer questions by category and shape. It is not a sample of what any one organization's customers ask, and wording effects on other kinds of prompts are not measured.
- How many wordings are enough. The panel holds two wordings per reworded pair and three specificity levels per intent. It cannot say how many wordings a tracked intent needs, or whether a third wording adds sources a second does not.
- Brand mentions. The unit is the cited domain. Whether an answer names a brand in its text, with or without a citation, is not measured here, so these figures are not a share of voice reading (see AI share of voice and Share of Citation for the metrics this page leaves aside).
- Which sources, or why. The ladder says how much cited volume two sets of runs share, not which domains drop out or what in the wording or the engine causes it. Nothing here says a placement or page caused a citation.
- Other engines and versions. Engines are observed through the models listed in the release manifest, which may differ from consumer apps, and Gemini and the two Google surfaces enter only the pooled rows.
- Other periods. The figures describe runs through October 20, 2026. The runs of a question cluster by date and engine (see run-date clustering), so halves are not independent samples, and the page reports the effect of the split rule rather than a single interval for the level.
For how a ranking depends on the questions behind it, see the question set behind a ranking. The Machine Relations Index publishes the release this page reads.
Method note. Release mri_score_v2.1+2026-10-20+bf82d9d98978; the release manifest at machinerelations.ai/data/mri-release-manifest.json carries the artifact hash, bf82d9d989784a62103e9bdaaaf71b33285a0e41100a8f853f8dbdbea4fa931b. Unit: a run, one question asked of one engine on one day; each question-engine-day cell holds one run. Window: runs from July 6 to October 20, 2026. Panel: the questions of the mri_v2 panel in the Index run ledger, excluding the news_topic shape, 980 questions in 24 categories. Only successful runs of scoring-eligible questions are used, and runs from engine-days the Index excludes for collection health are left out, as in the release. A cited source is the Index's resolved domain for a cited URL, counted once per run. Denominators: 22,620 runs (5,316 ChatGPT, 5,420 Claude, 5,023 Gemini, 817 Google AI Mode, 644 Google AI Overviews, 5,400 Perplexity), 16,975 runs with a citation, 135,864 domain-run pairs. These denominators can be re-read from the public artifact at machinerelations.ai/data/machine-relations-index.json: sum question_context.strata over the strata whose question_shape is not news_topic (question_count 980, the engine_runs values above, cited_source_runs 135,864), and read paraphrase_group_count (700 groups) and specificity_questions for the design. The overlap figures themselves need the record of which domains each run cited, which is held in the Index run ledger and is not in the public file, so they cannot be recomputed from the public artifact alone.
Computation. Split rule: the DuckDB hash (version 1.4.4) of the run identifier, modulo 2, assigns each run to a half within its question, so the headline split is deterministic and uses no random seed; comparisons need at least 8 domain-run pairs in each half; the rerun compares the first half with the second once; reworded comparisons use the 280 Index paraphrase groups, broader or narrower comparisons the 140 specificity ladders, both in both half orientations (first half of one question against second half of the other, and the reverse); further-apart comparisons use all qualifying pairs in both orientations, except that the different-category row uses three partners per question chosen by hash order. Shared volume is the sum over domains of the smaller of the two halves' shares of cited volume; Jaccard uses distinct domains. Matched pairs keep pairs whose two questions both have a rerun score, divide the mean pair overlap by the mean of the two questions' rerun levels, and take 95 percent intervals from 1,000 resamples of the 24 categories. Other splits: two other hash salts and one hash stratified by engine; alternating run dates (date parity counted from May 10); earlier against later runs (each question's runs divided at its median run date). Without-the-head runs drop the 25 domains ranked 1 to 25 in the release. The independent check recomputed the ladder from a run-level export in a separate program using five seeded equal-size random splits. Single-run pairs use every pair of cited runs of the same question. The bootstrap used a fixed linear congruential generator seeded 12345, and the independent program used seeds 1 to 5. Counts and shares were computed on October 20, 2026 from the run ledger for this release, whose denominators match the public artifact's question_context.
Frequently asked questions #
Do the cited sources change when a prompt is reworded? #
Somewhat. In a panel of 980 questions and 22,620 answer runs, a half of the runs of a reworded question shared 39.4 percent of cited-source volume with a half of the original's runs, against 46.1 percent between two halves of the original question itself. Depending on how the runs were divided, a reworded question kept 77 to 88 percent of the rerun overlap.
Do the cited sources change when a prompt is broadened or narrowed? #
Substantially. A broader or narrower version of the same intent shared 23.0 percent of cited-source volume, against 46.1 percent for a rerun, so it kept 44 to 49 percent of the rerun overlap. A question with a different intent in the same category shared 7.8 percent.
Does one run of a prompt show which sources AI engines cite for it? #
It shows a small sample. A cited run listed 8.0 domains on average, and two single runs of the same question shared 14.3 percent of their cited volume, 31.6 percent when the same engine ran both and 9.0 percent when the engines differed. A domain cited in one run of a question's first half was cited again in the other half 37.8 percent of the time, and one cited in five or more runs 95.6 percent.
How many prompts or wordings should be tracked to measure AI visibility? #
This analysis does not give a number. It supports tracking an intent over many runs, reading rewordings of a question as a mild change that behaves much like extra runs, and tracking broader or narrower versions as separate questions, with the run count stated beside every reading. It does not measure how many wordings an intent needs.