Your website says one closing time. A major listing says another. Neither page set out to mislead anyone. One of them is old, or was edited by someone who never told the other. You may have lived with that disagreement for months without knowing it exists. The assistant reading both pages has to answer tonight, in one sentence, and pick a version.
What happens when sources disagree
Research Note 003 was built to measure that choice. The design was narrow on purpose, so the pattern would be readable. Six long established businesses in Tokyo were selected. For each one, a single fact published on its official website was tested: a regular closing day, or a closing time. Three businesses were cases where a major listing disagreed with the official source. Three were cases where every listing checked agreed with it.
Four AI systems, ChatGPT, Gemini, Perplexity and Claude, were asked about that one fact, in Japanese and in English, three times each, through APIs. The total was 144 answers. Each answer was classified by whether, on that fact, it agreed with the official information, conflicted with it, or gave no single value for it. The classification made no judgement about which version was correct in the world. It measured agreement with the business’s own published information.
Where the sources agreed, the result was clean. Across the three agreement cases, 70 of 72 answers agreed with the official information and 0 conflicted with it. The remaining 2 gave no single value. No system asserted a value that differed from the official information.
Where the sources disagreed, the picture changed. Across the three disagreement cases, 58 of 72 answers agreed with the official information, 7 conflicted with it, and 7 gave no single value. The conflicts appeared only where the sources themselves disagreed.
Overall, of 144 answers, 128 agreed with the official information (89 percent), 7 conflicted, and 9 gave no single value.
What we measured. Research Note 003, measured November 8, 2026, asked four AI systems about one fact for each of six Tokyo businesses, in Japanese and English, three times each: 144 answers, collected via API with search enabled. This was a four system study with no Google AI leg. Agreement cases: 70 agreed, 0 conflicted, 2 gave no single value, out of 72. Disagreement cases: 58 agreed, 7 conflicted, 7 gave no single value, out of 72. Of the 7 conflicting answers, 5 matched the disagreeing listing’s value and 2 gave a third value found in neither source. Five of the 7 conflicting answers came from one system in the configuration tested.
The listing usually wins, and sometimes neither version wins
The 7 conflicting answers split into two kinds.
Five matched the disagreeing listing’s value exactly. That is the pattern to remember. When a website and a listing argue, the listing often carries more weight in the pile the machine reads: it arrives with a platform behind it, it is structured for machines, and it sits in more company. Your own website is one voice. The listing is one voice amplified.
Two gave a third value that appears in neither the official source nor the disagreeing listing. In one case, where the official information implied a weekly closing day without stating it in so many words, answers concluded that the business was open every day. When facts are unclear, the machine does not always choose between them. Sometimes it settles the argument by inventing a peace treaty. You cannot brief against improvisation. You can remove the ambiguity that invites it, by stating facts so plainly that no treaty is needed.
The 7 answers that gave no single value in the disagreement cases also matter. Six presented both versions without settling on one, and 1 gave a range spanning both values. Presenting both versions can read as a useful caution, but for a customer deciding on the strength of the answer alone, it leaves the question unsettled.
Concentration is worth stating carefully. Five of the 7 conflicting answers came from one system, in the specific model and search configuration tested, and that system also accounted for 5 of the 7 no single value answers in the disagreement cases. Another system agreed with the official information in all 18 of its disagreement case answers. The study describes the tested configurations on one day. It is not a ranking of AI systems.
Agreement is a rung, not a tidy up
This result is why Agreement sits near the bottom of the Trust Ladder, as rung two: The Trust Ladder: Five Rungs Between Your Business and an AI Recommendation.
The machine in this study behaved like a careful reader of a messy pile. When the pile agreed, it repeated the agreement. When the pile argued, it resolved the argument quietly, usually in favour of the louder source, and never reported that a fight had taken place. The errors owners blame on AI were, in this measurement, mostly errors published in two places at once. The goal is not to persuade an assistant to trust you. It is to remove the disagreement that gives it a choice. This is the rung you control most fully, and it needs a decided version of each fact and the persistence to make every copy match it.
Two cautions belong beside the numbers. First, conflict here means conflict with the official information, not proof that a listing is wrong. In one case the listing presented its value as a recent change, and the official website may not have caught up. The finding stands either way: the answer a customer receives can differ from the information the business itself publishes. Second, an answer matching a listing’s value does not prove the listing caused the answer. The study records agreement between values, not the path an answer took, and it does not confirm that aligning listings changes AI answers. What it shows is where the contradictions sat: only where the sources disagreed.
Step one: one source of truth
Every fix stands on one decision made before any correcting begins. For each fact, there is exactly one place where the true version lives, and every other place is a copy.
For most businesses that place should be your own website, the property you control and can change the day a fact changes. Pick the place you will actually keep correct, and write the choice down where anyone who helps you will find it. A source of truth held in one person’s memory has the same life expectancy as that memory.
Build it fact by fact before touching anything else: hours with holiday and seasonal exceptions, prices in the form customers pay them, location in the exact words a map can use, the offer, contact routes, and the policies customers ask about. Each fact written once, completely, in the place you chose.
Step two: the correction list
With a source of truth in place, disagreements stop being mysteries and become arithmetic. A correction list is a table with one row per place your facts appear: what that place says now, what the source of truth says, the gap between them field by field, the action, and the date the correction was made.
Build the inventory first, and expect it to be longer than you thought: your website and each location page, the major listings and booking platforms, guides and directories, social profiles, and old profiles under an old name or address. In our case screening for Research Note 003, one disagreeing listing turned out to be an old record from before a relocation, still standing and still being read. Ghost records like that compete with the real business quietly.
Work the list in order of influence. Correct the dominant sources from the citation map first, because the machine reads them more and weighs them more. Match the source of truth exactly, including the awkward fields owners skip: holiday hours, entrance notes, secondary phone numbers. Skipped fields are where disagreements retire to. The list is finished when there is nothing left for the machine to argue with.
One rule keeps the list honest. A correction list may only promise what can actually be executed. Some entries cannot be fixed: a listing controlled by someone else, a platform that will not accept your edit, a directory that no longer operates. Those entries belong in your records as findings, labelled as beyond reach, with the mitigation noted: strengthen the correct sources around them. Promise only what you can do. Report everything you found.
Step three: keep the record agreed
When a fact changes, change it in the source of truth first, the same day, then propagate it down the same list in the same order. A fact updated in one place only is a new disagreement, scheduled.
After corrections, re-measure with the same discipline that produced the list: the same stranger questions, repeated, in each language. Not every answer will change. What changes is the record the answers are built from, and the measured expectation is narrow: where the sources agree, the contradictions stop appearing.
Key takeaways
- In Research Note 003, agreement cases produced 0 conflicting answers out of 72. Disagreement cases produced 7 out of 72.
- Of the 7 conflicting answers, 5 matched the disagreeing listing’s value and 2 gave a third value found in neither source.
- Overall, 128 of 144 answers agreed with the official information, 7 conflicted, and 9 gave no single value.
- Five of the 7 conflicting answers came from one system in the configuration tested. The study is not a ranking of systems.
- Fixing starts with one source of truth, then a correction list worked in order of influence, with unfixable entries reported as findings.
- When a fact changes, update the source of truth first and propagate the change everywhere in the same week.
Find your disagreements before your customers do
The free Quick Scan checks public search results for disagreements in basic business details. If Quick Scan finds concrete errors, the Full Audit checks every factual claim in its answers against the facts you confirm, and the Fix Pack produces the corrections as ready to apply deliverables.
Check your business
See what public sources say about your business for free.
Run Quick ScanAbout Nixuma
NIXUMA is a GEO (Generative Engine Optimization) and AI-accuracy company. It measures what AI assistants — ChatGPT, Claude, Gemini, Perplexity and Google's AI — say about a business, checks those answers against the business's own confirmed facts, and publishes the measurements.