More
Free Blog Audit Start a Blog

AI Business Accuracy Report 2027: Study Protocol

Short answer: this page is the method for the AI Business Accuracy Report 2027, published before a single answer has been collected. We will ask ChatGPT, Google Gemini, Perplexity and Google AI Mode what 500 randomly chosen UK companies sell, what they charge, who they compete with and whether they offer a particular service, then check every factual claim against a record we build before we see any answer. Results are planned for June 2027.

Protocol v1.0, published 2 November 2026. Status: no data collected. Any change before collection will be dated in the change log at the foot of this page.

Why we are publishing the method first

Businesses now hear about AI answers getting their details wrong, and most of the numbers they hear come from companies that sell the fix. Some of those studies are large. Few say in advance what they will count, and in the biggest one we found, the brands checked their own facts.

Fixing the questions, the sample and the scoring rules in public before collecting anything means the result cannot be tuned to fit a story afterwards. If the answer is that AI describes businesses well, we will publish that. If a rule in this protocol turns out to be wrong, the change and its date will be visible below.

What the study asks

  1. Primary question. What share of the checkable factual claims each AI platform makes about a business are incorrect or outdated?
  2. Does accuracy differ by industry and by company size?
  3. How often does a platform describe the wrong business, or put a competitor in place of the one asked about?
  4. Asked whether a company offers a service it does not offer, how often does the platform say yes?
  5. How accurate are stated prices, and how often does a platform give a specific price for a company that publishes none?
  6. How stable are answers over six weeks, and do answers from the platforms’ developer interfaces match what a customer sees in the app?

We will also look at whether some company features go with higher accuracy, such as a Wikidata entry, a published price list or facts that agree across sources. Those results will be reported as associations, never as causes. We have set no expected direction for any result.

Which businesses

The sample frame is the free monthly company data file from Companies House, which lists every registered UK company with its status, industry codes and accounts category. We will draw 500 active companies with a working website: 50 in each of ten industries, split across three size bands.

IndustryExtra source for checking facts
Software and SaaSVendor pricing pages and our dated price logs
Web hosting and website platformsVendor pricing pages and our dated price logs
Online retail and direct-to-consumer brandsProduct, delivery and returns pages
Travel and hospitality groupsCAA ATOL register where relevant
Consumer finance and fintechFCA Financial Services Register
Education and training providersOfsted or Office for Students registers where relevant
Subscription consumer servicesPlan and tariff pages
Fitness and leisure operatorsMembership pricing pages
Multi-office professional servicesSRA, ICAEW or the relevant professional register
Manufacturing and industrial suppliersProduct catalogues

Size bands. Within each industry, about 17 micro and small, 17 medium and 16 large companies, using the Companies House accounts category as a stand-in for size. It is an imperfect stand-in and the report will say so.

How companies are picked. Each industry and size group is shuffled with a random seed. We take companies in that order, find each website by searching the registered name, and confirm it by a company number or registered address on the site. A company we cannot confirm is replaced by the next one in the list, and every rejection is logged with its reason. The seed and the date of the Companies House file will be added to this page before the first company is drawn.

Left out on purpose. Local trades, care providers and sole traders, which our sister site Whito already tests (see interests and related work). Also dormant companies, holding companies with no customer-facing site, and companies named after a living person, so the study handles no personal data.

What the sample represents. UK companies with a confirmable website in these ten industries. Large firms are deliberately over-represented so they can be compared with small ones, so any figure for all businesses will be shown both as counted and weighted to the size mix in the Companies House file.

Which AI platforms, and how answers are collected

ChatGPT, Google Gemini, Perplexity and Google AI Mode, each in its default setup on the day. For every answer we record the date and time, the model name the platform shows, whether it searched the web, every link it cited and the full text.

The platforms’ terms limit automated collection. OpenAI’s terms bar programmatic extraction of output except through its API. Google’s terms bar automated access that breaks the instructions on its pages. Perplexity’s terms bar scrapers and bots. So collection is split in two:

  • All 500 companies are asked through the official developer interfaces of ChatGPT, Gemini and Perplexity, with web search switched on.
  • A subsample of 100 companies (10 per industry) is asked by a person in the ordinary consumer apps of all four platforms, including Google AI Mode, which has no developer interface. Comparing the two routes on the same companies tells us whether developer-interface audits reflect what customers actually see. That comparison will be published as a finding.

Every prompt starts a fresh conversation, signed out where the platform allows it, otherwise with memory and personalisation off. All platforms are asked about a company within the same 48 hours, and within 14 days of its record being built.

The seven questions

Square brackets are filled from each company’s record. The company is named with its category and town, the way a careful customer would ask.

#QuestionWhat it tests
Q1What does [Company], the [category] company in [town], sell or do?Core description, missing information, wrong business
Q2How much does [Company] charge for [main product or service]?Price accuracy, and prices given where none is published
Q3Who are [Company]’s main competitors?Whether named competitors exist and work in the same market
Q4What makes [Company] different from its competitors?Whether the claimed differences have any source
Q5Does [Company] offer [service]?Half the companies get a service they offer. Half get a plausible one they do not, as a control
Q6What are [Company]’s strengths and weaknesses according to its customers?Reputation claims and whether they cite anything
Q7I need [requirement matching its core offer]. Would you recommend [Company], or someone else?Recommendation and competitor substitution

The control service in Q5 must be offered by at least two of the company’s competitors and be absent from its own site and site search. A second person checks each one before use. On the 100-company subsample, Q1 and Q2 are also asked in two other wordings, because our own paraphrase study found that rewording a question changes which sources an AI answer draws on.

The record each answer is checked against

For each company, a record is built before any AI answer is collected, by someone who has not seen an answer about that company. Every field carries its source link, the date it was read and a saved copy.

FieldSource
Legal and trading name, company number, status, incorporation date, accounts categoryCompanies House
What it sells, services offered and not offered, locations servedThe company’s own website and site search
Published prices for the main offer, with billing termsPricing pages, and our dated price logs for software and hosting
Regulatory statusThe relevant public register
Review profile: platform, rating, number of reviews, date readPublic review pages
Earlier versions of key factsThe Internet Archive, to tell outdated claims from invented ones

If a company’s own sources disagree, for example two different prices on two of its pages, the field is marked conflicted and claims about it are scored as unverifiable, not wrong. The record is the company’s public information. If a company’s own website is wrong, an AI that repeats it is scored correct, and the report will say that.

How answers are scored

Each answer is split into single checkable statements. “The company sells accounting software for small firms from £12 a month” is two: what it sells, and what it charges. Each statement gets one verdict.

VerdictRule
CorrectMatches the record
Partly correctRight in substance, wrong in a detail, such as the right product under the wrong plan name
OutdatedTrue in an archived version of the company’s pages, no longer true now
IncorrectContradicted by the record and never true in the archive
UnverifiableNothing confirms or contradicts it, or the company’s own sources conflict
OpinionNo attributed source, such as “great customer service”. Not scored as true or false. Scored for whether any source supports it

Prices count as correct within 10 percent of the published price for the same plan and billing term, partly correct within 25 percent, and incorrect beyond that. A specific price for a company that publishes none is counted separately as an unsupported price.

Flags on whole answers: wrong business described; a competitor put in place of the company asked about; yes or no given wrongly on the Q5 control; core offer missing from Q1; refusal or “I don’t know”; a cited link that does not support the statement it is attached to (checked on a random fifth of cited statements).

Primary measure: incorrect plus outdated statements, divided by all checkable statements, for each platform. Secondary measures: the share of companies with at least one incorrect or outdated statement; the outdated share of all errors; price accuracy; the false-yes rate on the control; wrong-business and substitution rates; how many named competitors are real and in the same market; unsupported differences; stability across repeats; and the gap between developer-interface and app answers.

Checks on our own work

  • Pilot. 30 companies outside the main sample, in January 2027, to test the questions and the scoring rules. Pilot data will not appear in the results.
  • Rules frozen after the pilot. Any later change is dated on this page.
  • Second coder. 75 of the 500 companies are scored again by a second coder who has not seen the first scoring. Disagreements are settled and logged. If agreement falls below a Cohen’s kappa of 0.75, the rules are revised and the affected statements re-scored.
  • Same window. All platforms are asked about a company within 48 hours.
  • Fact-check before publication. The draft report is checked against the data by someone who did not write it, as with our other studies (see How we test).

Analysis plan

Rates are reported with 95 percent intervals by platform, industry and size band. Because every company is asked on every platform, platforms are compared company by company. Statements cluster within companies, so intervals are computed by resampling companies, not statements.

With 500 companies, a share measured across the whole sample is precise to about plus or minus 4.4 percentage points at worst. Each industry of 50 is precise to about plus or minus 14 points, so we will only call an industry difference when the intervals do not overlap. Each size band of about 167 is precise to about plus or minus 7.6 points.

Repeat testing

A panel of 100 companies is asked Q1, Q2 and Q5 on all four platforms twice more, two weeks and six weeks after the first round. The same panel is planned for a six-month repeat in late 2027.

What this study will not tell you

  • It is one country, ten industries and one collection window in spring 2027. AI answers change, and every figure will be a measurement of that period.
  • It does not measure lost sales. A wrong answer is a risk to a business, and this study measures how often that risk appears, not what it costs.
  • Developer-interface answers may differ from what customers see. The 100-company comparison measures that gap. It cannot remove it.
  • Scoring long answers involves judgement. The written rules, the second coder and the published statement-level data are there so anyone can check ours.
  • Links between company features and accuracy are associations, not causes.

How we treat the businesses in the sample

The study measures what AI platforms say about businesses. It does not suggest that any business has done anything wrong. A company is named in the report only if it has been told what the AI said and what our record shows, and given ten working days to respond, and its response is printed beside the example. Where an AI answer makes a damaging false claim about a named company, we describe the type of error and do not repeat the claim word for word. Any individual named in an AI answer is removed from the published data.

Interests and related work

Blogging Titan and Whito share an owner, and both offer AI visibility services. Whito’s research covers AI answers about UK local trades and care providers, including an emergency plumber pilot, a home care register check and a re-ask panel. This study leaves those sectors out so the two do not measure the same thing.

Other published work on the same question includes Searchable’s AI Brand Misinformation Report, where brands marked AI-stated facts about themselves as true or false; Searchable’s test of 165 London businesses against Companies House and LinkedIn; Insites’ check of 10,000 US local businesses against Google Business Profiles; and Seer Interactive’s phone number test. This protocol adds an independent record built before any answer is seen, a random sample, a split between outdated and invented claims, and a control question.

Timeline

WhenStage
2 November 2026Protocol v1.0 published (this page)
December 2026Seed and Companies House file date added; sample drawn
January 2027Pilot on 30 companies; scoring rules frozen
February to March 2027Records built for 500 companies
March to April 2027AI answers collected
April to May 2027Scoring, second-coder check, analysis, right of reply
June 2027Report, PDF, slides and statement-level dataset published

Challenge the method

If you think a rule here is wrong, or that the study will miss something that matters, email contact@bloggingtitan.com before collection starts. Changes we accept will be logged below with the date and the reason. The full list of our studies is on the research page.

Change log

VersionDateChange
1.02 November 2026First publication. No data collected.

Cite this protocol

Blogging Titan (2026). AI Business Accuracy Report 2027: study protocol, v1.0. Published 2 November 2026. https://bloggingtitan.com/research/ai-business-accuracy-protocol/