An AI engine does not search for the question you asked. In the Machine Relations Index answer ledger, Claude's web searches for buying questions had a median of 7 words where the question had 16. Claude never searched the question as written (0 of 8,059 searches), and it put a year in at least one search in 40.4 percent of its searching runs, though no question in these runs contained a year. 2026 was in 26.1 percent of those runs, 2024 in 14.6 and 2025 in 8.0. On "best X" questions, 2026 was in half the runs (50.0 percent).
The year reaches the answer. Across 366 questions that Claude answered both ways, the sources it cited carried 2026 in their address more often in runs where its own search said 2026: 18.2 percent of cited URLs against 6.3, a gap of 9.2 points on average across the questions.
This page reads what the ledger recorded as each engine's search queries. It does not show what a person's own app would search, and it does not show that a page which carries a year is cited because of it.
What was counted #
- Population. Successful runs in the Index's buying-question panel (
mri_v2), 2026-07-06 to 2026-11-07, through the engines' APIs, runs that name a brand in the prompt excluded. The ledger records the search queries of three engines: Claude, ChatGPT and Gemini. For Perplexity, Google AI Mode and Google AI Overviews it holds none, so this page says nothing about them. - Claude. 7,726 runs, of which 4,622 searched the web (59.8 percent) on 890 questions, with 8,059 searches in all. Every one of them ran on
claude-haiku-4-5-20251001, the only Claude model in these runs. This page is about that model through its API, not about Claude's apps. - ChatGPT. 7,687 runs, 5,996 with a recorded search (78.0 percent), and 5,996 recorded searches.
- Gemini. 6,657 runs, 556 with recorded searches (8.4 percent) and 1,634 searches. The recorded share is small; whether the other runs searched is not known, so Gemini is a side note here.
- "Added" words. A word counts as added when a search contains it and the question that prompted it does not.
- Intervals. 95 percent intervals resample the questions with replacement, keeping every run of a question together (200 draws, seed 11).
The questions are the Index's buying questions, which come in the six shapes listed in the shape table below. They are library prompts, not measured buyer demand.
Claude searches for a short, different question #
| Engine | Runs that searched | Searches | Searches per run | Median words in the search | Median words in the question | Search is the question word for word | Search words found in the question |
|---|---|---|---|---|---|---|---|
| Claude | 4,622 | 8,059 | 1.74 | 7 | 16 | 0.0% | 60.1% |
| Gemini | 556 | 1,634 | 2.94 | 8 | 15 | 0.4% | 50.4% |
| ChatGPT | 5,996 | 5,996 | 1.00 | 16 | 16 | 100.0% | 100.0% |
Claude's searches average 1.74 per run, and 3,119 of the 4,622 runs (67.5 percent) ran two or more, so the second search is often a different angle on the same question. A question such as "is a dedicated procurement tool worth it or can we run purchasing through our erp" became "dedicated procurement tool vs ERP system 2024 2025" and "when to use standalone procurement software versus ERP".
ChatGPT's row is a limit of the record, not a finding about ChatGPT. The ledger's search field for ChatGPT repeats the prompt in all 5,996 runs, so either the API does not expose ChatGPT's rewritten search or the collector does not store it. The ledger cannot show whether ChatGPT rewrites.
What Claude adds #
Share of Claude's 8,059 searches that contain the word or phrase, and share that contain it when the question did not:
| In the search | Any | Added by Claude |
|---|---|---|
| vs or versus | 20.0% | 12.6% |
| compare, comparison or comparing | 12.4% | 11.1% |
| best | 14.3% | 7.8% |
| price, pricing or cost | 8.2% | 5.5% |
| review or reviews | 1.4% | 1.1% |
| a year (2024, 2025 or 2026) | 23.7% | 23.7% |
It turns open questions into comparisons. Gemini adds "best" (5.6 percent of its searches), "vs" (5.1), reviews (3.4) and, in 1.2 percent, the word Reddit, on the smaller recorded base.
The year #
No question in these runs names a year, so every year in a search was put there by the engine.
| Share of Claude's searching runs with the year in at least one search | Runs | Share (95% interval) |
|---|---|---|
| Any year | 4,622 | 40.4% (37.9 to 42.9) |
| 2026 | 4,622 | 26.1% (23.8 to 28.4) |
| 2024 | 4,622 | 14.6% (13.3 to 16.1) |
| 2025 | 4,622 | 8.0% (7.0 to 9.1) |
| 2024 or 2025 and not 2026 | 4,622 | 14.2% (12.9 to 15.6) |
In November 2026, a seventh of Claude's searching runs asked only for 2024 or 2025. That looks like the model's sense of the date lagging the calendar; the ledger does not show why.
The year depends on the question. By shape:
| Question shape | Questions | Runs | Any year | 2026 | 2024 or 2025 |
|---|---|---|---|---|---|
| Best X | 165 | 936 | 65.5% (59.1 to 71.3) | 50.0% (44.3 to 55.9) | 16.3% (12.3 to 19.3) |
| Top list | 167 | 970 | 50.6% (45.1 to 56.6) | 32.7% (28.4 to 37.6) | 18.5% (15.1 to 21.9) |
| X versus Y | 166 | 899 | 35.4% (28.9 to 40.8) | 20.7% (15.5 to 25.9) | 14.7% (11.0 to 18.4) |
| Is X worth it | 162 | 828 | 33.1% (27.4 to 39.1) | 16.5% (12.8 to 20.6) | 16.8% (13.0 to 20.0) |
| Problem first | 70 | 227 | 30.0% (21.2 to 39.8) | 18.9% (11.0 to 27.3) | 11.0% (6.7 to 16.6) |
| How do I choose | 160 | 762 | 13.4% (10.2 to 17.1) | 7.2% (4.2 to 9.6) | 6.4% (4.3 to 8.6) |
It also depends on the topic. 2026 was in 42.8 percent of Claude's searching runs on family-software questions and 35.2 on HR and talent, against 9.5 on local services and 3.0 on industrial categories with few reviews.
The year shows up in what Claude cites #
For each question that Claude answered both with and without the year in its own search, the table compares the share of cited URLs that carry that year in the address, with the year in the search and without it. Only runs with at least one cited URL count.
| Year | Cited URLs with the year in the address, runs that searched it | Same, runs that did not | Questions compared | Mean gap across questions (95% interval) |
|---|---|---|---|---|
| 2026 | 18.2% (1,228 of 6,733) | 6.3% (1,194 of 19,103) | 366 | +9.2 points (7.5 to 10.6) |
| 2025 | 5.1% (107 of 2,106) | 1.0% (227 of 23,730) | 234 | +3.2 points (2.1 to 4.4) |
| 2024 | 1.3% (51 of 3,872) | 0.3% (57 of 21,964) | 353 | +1.3 points (0.6 to 2.0) |
The interval excludes zero for every year, and the gap is small for the older years: an engine that asks for 2024 finds a 2024 address on 1.3 percent of what it cites. It is also what a search for "2026" should do, since many URLs and titles carry their year. The comparison does not show that the year changes which sources are chosen, only that a stamped search lands on stamped addresses more often. The year-in-address study shows that 5.5 percent of all cited addresses carry 2026.
What this changes #
- Test the search, not only the question. A prompt-tracking report that records one wording of a question is measuring a different thing from what Claude searches. In this ledger Claude's search is less than half the question's length, has a comparison word the question did not, and often has a year.
- Put the current year on pages that answer "best" and top-list questions. Half of Claude's runs on "best X" questions searched 2026. A page whose title and address say 2026 matches the words it sends; this page does not show that it earns the citation.
- Date pages honestly and keep them current. One run in seven searched only for 2024 or 2025, so an older year in a title still matches some searches, but this page does not measure which year a reader or an engine prefers. Say on the page when it was last checked.
- Do not generalize to ChatGPT, Perplexity or the apps. The ledger holds no rewritten searches for them, and Claude's API model is the one measured here.
What this page does not show #
- Claude's apps. The model is Claude Haiku 4.5 through the API with web search. The apps may search differently.
- A cause. A run that stamped 2026 may differ from one that did not in ways beyond the stamp. The gaps are within the same question, not an experiment.
- Other engines. ChatGPT's recorded search is its prompt. Gemini's is recorded in 8.4 percent of runs. Perplexity and the two Google engines record none.
- Date effects. Runs of one question were not matched by date. The share of "best X" runs with 2026 was 55.7 percent in September (201 runs), 49.3 in October (639) and 42.7 in November (96); this page does not say whether that is a trend.
- Demand. The questions are library prompts, not buyer searches.
Method note #
Source: the Index answer ledger's run and citation-events tables, daily export of 2026-11-11 08:10 UTC, read with DuckDB and Python. Filters: the buying-question panel (mri_v2), runs not flagged as brand-prompted, successful runs with a non-empty recorded search_queries list, run dates 2026-07-06 to 2026-11-07, engines Claude, ChatGPT and Gemini. Totals for the share of runs that searched use the same filters without the search condition. A year is a four-digit 20xx token in a search; 2026, 2025 and 2024 are matched as strings. "Added" words are matched as whole words by regular expression. Cited-URL year matches look for the year as a whole number inside the URL string. The paired comparison pools events within each question for the runs with the year in a search and the runs without it, keeps questions with cited URLs on both sides, and resamples those questions (300 draws, seed 5). Question shapes come from the Index query library. No website is named in any table.
Frequently asked questions #
What do AI engines search for when they answer a question? #
Not the question. In the Index ledger Claude's searches had a median of 7 words against 16 in the question, none was the question word for word, and 1.74 searches ran per searching run. ChatGPT's ledger field holds the prompt itself, so its rewriting is not visible there.
Does Claude add the year to its searches? #
Yes, though no question in these runs contained a year. A year was in at least one search in 40.4 percent of runs: 2026 in 26.1 percent, 2024 in 14.6 and 2025 in 8.0. On "best X" questions 2026 was in half the runs.
Does Claude ask for old years? #
Sometimes. In 14.2 percent of runs, Claude's searches asked for 2024 or 2025 and not 2026, in November 2026.
Does putting the year in a page's address help it get cited? #
This page cannot say. Cited URLs carried 2026 in their address in 18.2 percent of runs where Claude searched 2026 and 6.3 percent where it did not, but the data does not show the year caused a citation.
Is query fan-out the same on every engine? #
The ledger cannot say. It records searches for Claude, ChatGPT and Gemini only, and ChatGPT's record repeats the prompt.
Why test a question's wording and the engine's search separately? #
Because they differ. On average 60.1 percent of the words in a Claude search came from the question; the rest were added, often comparison words and years. A report that scores one wording does not score the search.
See also: when Claude and ChatGPT search for buying questions, whether rewording a prompt changes cited sources and the year in cited addresses.