makefield

For PR and comms teams

Ask the AI about your organisation: a toolkit for PR teams

Question sets to copy, a log to keep and a plain guide to reading what AI assistants say about your organisation, your releases and your spokespeople.

Denes Csaszar · Makefield · Berlin

Start here

Back to AI answers and PR. This website sells nothing.

This page helps your team find out for yourselves what AI assistants say about your organisation, and how far to trust it.

Pick a level:

Your first result will probably be mixed and a little odd. That is normal. One answer is one draw, not a measurement.

Before you ask

Pick one purpose. For example: does a journalist get our facts right? Are we named when a buyer asks about our market?

Write down your own source first, and date it. Note what a correct answer contains: what you do, who leads you, your spokesperson's title, the main claim of your latest release. Ask the people who own those facts to confirm them. Without this, "accurate" only means "sounds right". Audit your AI visibility in an afternoon shows more.

Keep it safe

  • Ask about public facts only.
  • Do not paste unreleased releases, embargoed or client material, or personal data into an assistant.
  • Use a work account if your organisation requires it.
  • Before you ask about a named spokesperson, tell them, and check with your data protection contact that your organisation allows staff names in the assistant you use. Do not add private details. Keep the log only as long as you need it.
  • A log can also show that your team saw an answer on a given date. Ask your own counsel how long to keep it and who may see it.

The question sets

Six sets, one per purpose. Put your own names in the brackets. The IDs let a colleague ask the same thing next month.

One rule for every set: after each answer, ask in the same chat: "Where does this come from? Give me the links." It shows which sources the engine names. Open each link: some may not be the ones it used, and some may not exist.

W · Who you are

  1. What does [Organisation] do, and who is it for?
    What this shows: the plain description a stranger gets.
  2. Who leads [Organisation], and who speaks for it?
    What this shows: whether names and job titles are current.
  3. Where can I read more about [Organisation] for an article?
    What this shows: which pages the engine treats as the source.

R · Releases and news

  1. What did [Organisation] announce about [Release headline]?
    What this shows: whether the release is found, and whether numbers and dates stay intact.
  2. What has [Organisation] announced lately?
    What this shows: what the engine treats as news, and which source it prefers.
  3. Where can I find the original statement from [Organisation] about [Release headline]?
    What this shows: whether it points to your press page or to a copy elsewhere.

S · Spokespeople

  1. Who is [Spokesperson]?
    What this shows: wrong titles, old roles, or a different person with the same name.
  2. Who is an expert on [category] in [country]?
    What this shows: whether your spokesperson is found when nobody names them.
  3. What has [Spokesperson] said about [Release headline]?
    What this shows: invented, shortened or wrongly credited quotes. Check them against your approved text.

M · Your market

Marketing and sales teams can use this set as it is: it shows what an assistant tells a buyer who asks about your field.

  1. Which [category] suppliers in [country] would you shortlist?
    What this shows: whether you are named without being asked about, and who appears instead.
  2. What is the best [category] for [need] in [country]?
    What this shows: the choice a buyer is offered.
  3. What are the alternatives to [Product]?
    What this shows: who stands next to you.
  4. [Organisation] or [Competitor] for [need]?
    What this shows: which points the engine uses to compare you.

P · Your reputation

  1. What are the criticisms of [Organisation]?
    What this shows: the story you did not write.
  2. Has [Organisation] had any scandals, complaints or legal disputes?
    What this shows: which old issue comes up first, and from which source.
  3. Has [Organisation] had a problem that it has since fixed?
    What this shows: whether the end of an old story is on record, or only its headline.
  4. Is there a problem with [Product]? Should I avoid it?
    What this shows: how a worried asker is answered.

Answers here and in M4 may repeat claims about real people or other companies. Keep them in the log, check them, and do not share them outside the team.

F · One fact at a time

  1. What is [Organisation]'s [one checkable fact, for example founding year or head office]?
    What this shows: one fact, asked openly and checked against your own source. Do not put the answer in the question.
  2. What does [Product] cost, and what is included?
    What this shows: invented prices and old specifications.

How to ask

  1. Ask as a stranger would. Signed out, or memory and custom instructions off, in a fresh chat. Why this matters.
  2. Use the asker's words, not yours. No slogans, no "leading" or "innovative".
  3. Leave your name out of market questions. Set M is about whether you are named at all.
  4. One question per chat. The only follow-up is the question about sources.
  5. Never type "search the web". Use the engine's own search switch, and note it. Why.
  6. Freeze the wording. A changed word is a new question.

Company voice: "Why is [Organisation] the leading innovative partner for [category]?"

Buyer voice: "Which [category] suppliers in [country] would you shortlist?"

Run it and log it

  1. Who: one person asks, a second reads the log.
  2. Freeze 5 to 10 questions from the sets, with their IDs.
  3. Fix the conditions: signed out or memory off, fresh chat, search set with the engine's own switch.
  4. Pick three engines your audience uses. Use the same three every time.
  5. Ask each question two or three times, each in a new chat.
  6. Log straight away, before anyone discusses the answers.
  7. Repeat on another day, then monthly and after a release, rebrand or crisis. Decide beforehand what would count as a real change.

The cost, said plainly: 6 questions × 3 engines × 3 asks = 54 answers, about 90 minutes to ask and save. Checking the facts takes extra time.

Log fields, one row per answer: date and time · engine and mode · signed in or out, memory, search · language and country · question ID · ask number · full answer saved · named without being asked (yes, no, or n/a when the question names you) · each fact right, wrong, outdated or can't tell, against your own source · cited links, and whose they are (who AI cites) · who else is named · notes and initials.

Paste this line into the first row of a spreadsheet:

Date	Time	Engine	Mode	Signed in	Memory	Search	Language	Country	Q-ID	Ask	Answer file	Named	Facts	Links	Others	Notes	Initials

Preserve first, interpret later. Take a screenshot with the time visible, and note the address and date of each cited page. Keep copies for your team's internal record only, if your organisation's rules allow it. Do not republish them. Never edit the original; work on a copy.

Read one answer

Ask four questions of each answer: are you named? Is it right? Whose sources? Who appears instead? How to read your own AI-visibility audit explains each one.

What this check cannot tell you

A log shows: we saw this, on this date, under these settings. It does not prove what an engine says to everyone. It says nothing about languages, countries or engines you did not test. It cannot tell you how often buyers see these answers, whether your work caused a change, or what comes next month.

Do not conclude:

Worked example

Fictional example: a made-up company, invented answers.

"Sample Tiles" (a made-up company) sells floor tiles to builders. Its team wrote down its facts, then asked sets W and M.

What the team concluded: the wrong year appeared in 1 of 2 asks, next to a directory page. They will open that page, check it against their own source, and ask again next week.

What they did not conclude: "AI thinks we were founded in the wrong year", or "we are invisible in our market". One market answer is one draw.

Why one answer is not enough

In Makefield's floor test, answers to the same question from different rounds were less alike than repeats within one round in 85 of 103 pairs. It compared wording, not facts.

The IAB puts it plainly (Measuring Visibility in the AI Era, August 2026, page 32): "Single-response measurement is not measurement." One AI question, two answers is a five-minute test you can try now.

Calculation, not a Makefield measurement

How far to trust your own log

Count how often you asked one question, and how many of those answers named you or got the fact right. The range shows what the true share could plausibly be.

Same share seen (60%), more asks: the plausible range (95%) narrows
Your logPlausible range
3 of 5about 23% to 88%
6 of 10about 31% to 83%
12 of 20about 39% to 78%

Calculation, not a Makefield measurement. It assumes each ask is independent; real answers can vary more.

The formula

This is the Wilson score interval at 95%. With k answers out of n asks, the share you saw is p = k / n, and z = 1.96.

centre = (p + z² / 2n) / (1 + z² / n)

half-width = z / (1 + z² / n) × √( p(1 − p) / n + z² / 4n² )

The range runs from centre minus half-width to centre plus half-width. Unlike the simpler "plus or minus" formula, it stays between 0% and 100% and still gives a sensible range when you saw 0 or all of the answers.

If an answer is wrong

Do not argue with the chat. Confirm it across repeats, find the page it reads, and fix your own sources first: AI got a fact about you wrong.

If a wrong answer is about a person, such as a named spokesperson, tell them first.

What to tell your boss

"We asked 6 questions, 3 times each, in 3 assistants. Here is what we saw, and what we do not yet know."

Sources and evidence

For context: in 2025, 32% of people aged 16 to 74 in Germany, 39% in Austria, 30% in Hungary and 47% in Switzerland used generative AI tools in the previous three months (Eurostat, official statistics, dataset updated June 2026). Across 48 markets, 10% use AI chatbots for news each week (Reuters Institute Digital News Report 2026). Limit: general use and news use, not research about companies.

How to cite: Denes Csaszar, "Ask the AI about your organisation: a toolkit for PR teams", Makefield, makefield.io/pr/kit, last checked 26 October 2026, version 1. Link one set as makefield.io/pr/kit#set-m.

Last checked 26 October 2026. What changed: version 1, first published 28 October 2026.

Makefield is an independent research project on AI answer accuracy, based in Berlin, Germany. This website sells nothing.

Also for PR teams: Measure what AI does with a press release · Five questions to ask of an AI-visibility report