For PR and comms teams
How to measure what AI does with your press release
Ask the same questions before and after one release, repeat them, and write down what changed. A plan your team can run itself, and what it can and cannot show.
This page is a method, not a result. It reports no measurement, and Makefield does not run it for you. This website sells nothing. Back to AI answers and PR.
What a reading can show
- Which facts from your release turn up in AI answers, right, wrong or missing.
- Which sources the answers cite: your own page, the outlets that carried the release, or others.
- Whether that moved after the release by more than the normal spread between repeats and between sittings.
What it cannot show
- That the release caused a change. Other news and engine updates happen at the same time.
- Reach, money or return. It does not rank outlets.
- What other people were shown. Account, location and settings change answers.
What it costs
Full run: 8 questions × 2 engines × 3 asks × 4 sittings = 192 answers. Plan for about 9 to 10 hours in all, spread over a month.
Minimum run: 6 questions × 1 engine × 2 asks × 4 sittings = 48 answers, about two and a half hours. It is rougher, so read it more carefully.
Hours are our estimate for asking, saving and reading. Asking in two languages doubles the count. Start well before the release: the first two sittings happen before it goes out.
The plan, step by step
1. Write the plan before the release, and freeze it
Take the release and list 5 facts in it that anyone can check: a product name, a figure, a date, a place, a name. Copy each one from your own published text. Add the questions, the engines, the language, the dates of the sittings, and what will count as a real change. Run the two pre-release sittings on different days. Date the plan and do not edit it afterwards. Without this, any difference you find can be explained after the fact.
2. Write three kinds of question
- Named: What has [Organisation] announced recently about [topic]?
- Named: What does [Organisation] offer in [category], and what does it cost?
- Named: Where can I find official statements from [Organisation] about [topic]?
- Buyer, without your name: Which [category] suppliers in [country] would you shortlist for [need]?
- Buyer, without your name: What is new in [category] this year?
- Buyer, without your name: What are the options for [need] in [country]?
- News: What has [Organisation] announced lately?
- News: What is the latest news about [Product]?
[topic] is the subject area, never the headline or an unreleased product name. Use identical wording in every sitting. For the minimum run, use two questions of each kind.
Ask them as a stranger would, in the words a buyer or journalist uses. The rules are on the toolkit page. Ask in the language your audience uses; in Hungary, ask in Hungarian and English.
3. Fix the settings
Same engines, same language, same location, signed out or memory off, a fresh chat for every ask. Write the settings down. Why signed out.
4. Take two sittings before the release
Ask every question at least three times in each sitting (twice in the minimum run). The spread between those repeats, and between your two pre-release sittings, is the normal noise. Use the larger. A change inside it is not a change.
One before and one after is not a baseline. The IAB puts the general point in one line: "Single-response measurement is not measurement" (August 2026, page 32).
5. Send the release
Note the exact time and every outlet that carries it.
6. Repeat the sitting after one week and after three weeks
The same questions, engines and settings. Do not change the plan.
7. Read every answer against your list
Is the organisation named? Which release facts appear, and are they right, wrong or missing? Which sources are cited, and how old are they? Who else is named?
8. Write the findings sheet
What was seen, under which settings, against which source. No opinion. Write counts as "x of n", never as percentages: "4 of 6 answers", not "67%". If one wrong fact keeps coming back, add a short note: what it was, where it seems to come from, and which of your own pages to check first. Then fix your own sources.
The sheet to copy
Paste this line into the first row of a spreadsheet, one row per answer:
Sitting Date Time Engine Mode Signed in Language Country Q-ID Ask Answer file Named Fact 1 Fact 2 Fact 3 Fact 4 Fact 5 Sources cited Source dates Others named Notes Initials
For each fact, write: right, wrong, missing, or can't tell. Named: yes, no, or n/a when the question names you.
Answers can repeat claims about real people or other companies. Keep answer files for your team's internal record only, and do not republish them.
Worked example
Fictional example: a made-up company, a made-up release, invented answers.
"Sample Tiles" (a made-up company) announces a new frost-proof tile on 3 March. Its plan lists five facts: the product name, the price, the launch date, the factory town and the frost rating. It uses the minimum run: six questions, one engine, two asks, four sittings.
- Two sittings before the release: the named questions return the old range only. 0 of 12 answers name the new tile in each sitting, so the spread was zero and anything above 0 is outside it. A zero spread from 12 answers is thin evidence.
- One week after: 5 of 12 answers name the tile. The price is right in 3 of those 5; the frost rating is wrong in 2, and both of those cite the same trade blog.
- Three weeks after: 6 of 12 answers name the tile. The buyer questions still do not name Sample Tiles.
What the team concluded: the release facts now appear in some answers; the wrong frost rating seems to come from one blog post, so they check that post and their own product page first.
What they did not conclude: "the release worked", or "AI knows our new tile". 12 answers per sitting from one engine are a small sample, and the change cannot be put down to the release alone.
Read it without blame
A reading describes what the engines did. It does not grade the release, the agency or the team. In one vendor's log data, press releases were about 1.1% of cited links (Muck Rack, May 2026; markets and prompts not stated). "Our release was not cited" is a common result, not a verdict.
For your boss
"We wrote our questions and facts down before the release, asked the same questions before and after, and counted. Here is what changed, and what we cannot say from this."
Keep it safe
- Never paste the release into an assistant before it is public. Ask only about published facts.
- If your release makes claims about products, prices or performance, your legal team may want to know you are checking what AI answers say about them. Ask them how long to keep the sheet.
- Keep the sheet inside the team. More on keeping it safe.
The method behind this
This page is the method in full. It is the same approach Makefield uses in its research: the plan is written and dated before the first question, each question is asked more than once, and the normal spread is measured before any change is read. The steps are free to use and change.
This is a research method, not an offer. Makefield sells nothing, takes no commissions and does not run this for anyone.
Sources
- IAB, "Measuring Visibility in the AI Era", August 2026, page 32.
- Muck Rack, "What Is AI Reading?", May 2026. Vendor log data; Muck Rack sells PR software. Prompt count and dates are not stated.
- 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.
How to cite
Denes Csaszar, "How to measure what AI does with your press release", Makefield, makefield.io/pr/press-release, last checked 26 October 2026, version 1.
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: Ask the AI about your organisation · Question an AI-visibility report