gpt-6.1-sol with search gave 13.7% of 444 software firms a founding year unlike Wikidata's

Earlier checks of what AI assistants say about companies used one brand, a few hundred, or a platform's own users. Seer Interactive tracked answers about itself (28,123 answers across six AI platforms) and found 95.9% of answers accurate when the platforms answered, with 31.2% of prompts declined. Searchable tested 165 London businesses, founding year among the facts, and found a 5% fabrication rate for small firms against 2% for large ones. Reputation Age and Bubbling checked the half-year results of 39 CAC 40 companies in ChatGPT and Gemini and found 40% of figures wrong. LLM Listed asked four assistants about 250 companies and found the headquarters right in 85% of responses. Timmermans Media assessed 1,917 API answers about 37 Dutch brands, founding year among the stable facts, and found 1.1% to 9.2% wrong by model. maxonline asked the ChatGPT app about 150 mid-sized companies in Germany, Austria and Switzerland in March and April 2026 and found 78% of the founding years it gave did not match the company records, on a sample with no stated selection method and without comparing search on and off. Searchable had about 1,700 brands that use its platform grade 32,556 facts that AI engines stated about them, and they marked 6.9% of history and founding facts false; the brands chose to take part, and engines were not compared with and without search. Academic benchmarks such as FACT-BENCH test models on general Wikidata facts, and PopQA (Mallen and others, 2023) found retrieval helps most on less popular entities, and Head-to-Tail (Sun and others, 2023) found models know far less about less popular entities.

This study asks two facts about 444 software companies drawn at random from an open public list, Wikidata, and checks the answers against Wikidata's own record. It asks every company four times: two OpenAI models, each from memory and with web search, so the effect of search can be read on the same companies. Those two things, a random draw from an open list and the same companies asked with and without search, are what this study adds to maxonline's and Searchable's. Its comparable figure, the share of stated years that differ from the key, is 17.1% from memory and 15.1% with search; the two studies differ in sample, region, interface and answer key, so the gap between their rates is not a measured effect. It also reads each company's home page for the founding date it states in its own markup, which shows how often the answer key and the company disagree. The per-company data is public.

With web search, 13.7% of companies got a founding year that differs from Wikidata's

gpt-6.1-sol, the first model in DataForSEO's list of OpenAI models on 3 November 2026, answered with web search on. For 61 of the 444 companies (13.7%, 95% interval 10.8 to 17.3) the year it gave differs from Wikidata's. It gave no year for 40 (9.0%) and the same year for 343 (77.3%).

Half the misses were one year out: the median miss is 1 year, and 31 of the 61 were exactly one year off. Counting a year within one as right, the share falls to 6.8% (95% interval 4.8 to 9.5), so the headline does not hold under that looser reading. 18 of the 61 misses were five or more years out.

The answer key can be wrong too. On the 129 companies whose home page states a founding date that agrees with Wikidata's, two sources that agree, the share of companies given a different year with web search was 10.1% (95% interval 6.0 to 16.5).

Across all four settings, 126 of the 444 companies (28.4%) got a different year at least once, and 20 (4.5%) got one in every setting. The same company can get a different year depending on the model, on whether it searches, or on the call: each setting was asked once.

For gpt-6.1-sol, web search filled gaps but did not lower the share of companies given a different year

From memory, gpt-6.1-sol declined to give a year for 186 companies (41.9%) and gave a different one for 44 (9.9%). With web search, it declined for 40 (9.0%) and gave a different one for 61 (13.7%). Search did not lower the share of companies given a different year. Among the years it did give, the share that differed went from 17.1% (44 of 258) from memory to 15.1% (61 of 404) with search: search answered far more often, and its answers differed about as often. The planned test of the share of companies, which treats the two readings as separate groups, finds no significant gap (adjusted p = 0.618). Found after review: compared company by company, 35 companies got a different year only with search and 18 only from memory, so search added 3.8 points (95% interval 0.6 to 7.0); p = 0.027 before correction and 0.243 after Holm correction across the nine comparisons.

Company by company, search turned 118 blanks into the right year, fixed 17 years that were different from memory and changed 6 right years into different ones. Found after review: it also turned 29 blanks into a different year, the largest source of its misses. 26 companies got a different year both ways, 23 of them the same different year from memory and from search.

The smaller gpt-5.4-mini missed more often when it answered. From memory it gave no year for 50.2% and a different year for 16.4%. When it did state a year, that year differed from Wikidata's 33.0% of the time (95% interval 27.2 to 39.5), against 17.1% (13.0 to 22.1) for gpt-6.1-sol. With web search, gpt-5.4-mini gave a different year to 14.9% of companies, close to gpt-6.1-sol's 13.7%; for the mini model, unlike the larger one, that is lower than from memory (16.4%), though not significantly.

Having the company's own site in the search results did not go with fewer misses

With web search, gpt-6.1-sol's search results included a page on the company's own website for 341 of the 444 companies (76.8%). For those companies, 14.1% of answers gave a different year; for the 103 whose site was not among the results, 12.6%. Having the company's own pages among the results did not go with fewer misses: found after review, the gap is 1.5 points with a 95% interval of -7.0 to 8.0.

One possible reason is that the company's sources disagree with each other. We read 377 of the 444 home pages; 146 of them (38.7%) state a founding date in Organization markup. That date agrees with Wikidata's year 88.4% of the time and is within a year 95.2% of the time. Of the 21 web-search answers that missed Wikidata's year for a company whose site states a date, 8 gave the site's date.

Stating a founding date in markup did not go with fewer misses: 14.4% of companies whose home page states one got a different year with web search, against 13.4% of the rest, whose home page states no date or could not be read (no significant difference; the gap could be -5.5 to 8.4 points).

Newer companies get blanks from memory

By founding decade, gpt-6.1-sol from memory gave no year for 88.3% of companies founded in the 2020s (143 of 162, 95% interval 82.4 to 92.4), 27.1% of those founded in the 2010s (29 of 107, 19.6 to 36.2) and 2.5% of those founded before 1990 (1 of 40, 0.4 to 12.9). By record age the pattern is the same: Wikidata records created from 28 August 2024 on (item Q130000000 and later) make up 249 of the 444 companies, and from memory gpt-6.1-sol gave no year for 69.5% of them, against 6.7% of companies with older records. With web search, it gave a different year for 16.5% of newer records and 10.3% of older ones; the gap could be -0.3 to 12.4 points and is not significant after correction (adjusted p = 0.533).

Most of the other groups have few companies. Of these cuts, only listing and employee band were among the planned tests, and neither differs significantly from the rest; industry and region are reported with their intervals and were not tested. By industry, with web search: software as a service 29.8% (14 of 47, 95% interval 18.7 to 44.0, not tested), computer security 14.8% (8 of 54), software industry 11.9% (39 of 329), cloud computing 0 of 14. By region: Europe 16.9% (29 of 172), United States 15.0% (27 of 180), Asia 1.9% (1 of 52, interval 0.3 to 10.1). Listed companies 12.5% (4 of 32), others 13.8%. Wikidata has an employee count for only 56 of the 444, too few for a size comparison; companies with under 200 employees got a different year 18.8% of the time (6 of 32).

Country: fewer misses, and a looser answer key

The headquarters country differed from Wikidata's country less often than the year. With web search, gpt-6.1-sol named a different country for 13 companies (2.9%) and none for 22 (5.0%). From memory, it named a different one for 23 (5.2%), and gpt-5.4-mini for 28 (6.3%).

Wikidata's country field records the country a company belongs to, which can be where it was founded or registered rather than where its head office is today, and for one company in the sample it holds the value "worldwide". A different country is a disagreement with that field: some of the different countries may be right answers, and an answer that repeats a wrong Wikidata country counts as matching.

What to do about it

• Ask the question both ways. 28.4% of the companies got a different founding year in at least one of four settings, so a single check of what an OpenAI model says about you is not enough. Ask from memory and with web search, on more than one model.
• Do not count on search to correct the record. With search, gpt-6.1-sol left 9.0% of companies without a year, down from 41.9% from memory, but the share of companies given a different year did not fall (9.9% from memory, 13.7% with search).
• Make every public source state the same date. Where a home page states a founding date, it agrees with Wikidata only 88.4% of the time, and of the 21 misses for companies whose site states a date, 8 matched the site's date. Pick one date, the one you want quoted, and put it on your site, your Wikidata item, and your company profiles. This study did not test whether aligning them lowers misses.
• Markup alone is not the fix. Companies with a founding date in Organization markup got a different year 14.4% of the time, the rest 13.4%. Treat markup as one of the sources to align, not as a switch.
• If you are new, expect blanks from memory. For companies whose Wikidata record dates from August 2024 or later, gpt-6.1-sol gave no year from memory 69.5% of the time. Make sure the pages that rank for your name state the date you want quoted.