How we measure AI visibility

In short:

• We ask AI assistants the same fixed homeowner questions for your trade, and repeat them every month.
• Every question names Singapore and is asked 3 times, signed out, in a fresh session.
• We report how often your firm is named ("named in X of 3 runs") and your share of voice against the other firms named.
• A separate readiness score covers what you have in place on and off your website. We never combine the two, and we never promise rankings.

The prompt sets

9 trades, 112 prompts

Source: weCDEF prompt sets, version 1.1 (18 Oct 2026), counted 25 Oct 2026.

The kinds of question we ask

Each prompt is written the way homeowners type, lowercase and casual, and is tagged with one intent:

• best: "best X" or "recommend an X";
• near-area: X near a named part of Singapore;
• cost: what a job costs;
• trust: licences, registration, deposits, warranties, red flags;
• comparison: X vs Y, DIY vs hiring, one platform vs another;
• problem: a symptom, and who to call;
• process: approvals, timelines, how often.

Every prompt names Singapore, and prompt IDs never change

Our checks run from a Singapore connection, and every prompt also names Singapore. No prompt names a client or a competitor. Each prompt has a fixed ID (for example WP-01), and we never edit or renumber a prompt once it's in use, because the monthly comparison depends on asking exactly the same question. New questions are added under new IDs at a quarterly review.

For a client, we may add 1–3 area questions for the parts of Singapore you really serve, copied from the standard ones with only the area changed. They're logged under their own IDs. The list is fixed before the first run.

How each run works

Signed out, fresh session, exact prompt text

Each run is a new, signed-out session. The prompt is pasted exactly as written, with no follow-up questions in the run we score. We never type the firm's name or website before the prompt. Every question is asked one at a time in an ordinary browser window, signed out, and the answer is screenshotted at the time. We don't use scrapers, APIs or bulk tools on the AI tools.

Where each tool is checked from

Client audits check ChatGPT and Google AI Overviews. Google doesn't show an AI Overview for every search, so when none appears we record "No AI Overview shown" for that run. That's its own outcome, not a miss. If Google blocks a search, we record "Blocked"; it isn't counted as a miss. Our published studies used a different setup: they checked ChatGPT and Perplexity, and Perplexity isn't part of our client audits.

A screenshot of every run, saved at the time

We screenshot the full answer and its list of sources before closing the window. Signed-out ChatGPT chats can't be reopened later, so a run without a screenshot is logged but doesn't count. For each run we log the tool, the model name if one is shown, whether web sources were shown, the date and time in Singapore time, the firms named and the sources cited.

Why 3 runs per prompt

Because the answers move. In our first pass on 18 Oct 2026, we asked one waterproofing question three times. Two firms were named in all three answers, but the third firm changed every time.

Source: weCDEF run notes for one waterproofing prompt, 18 Oct 2026 (from notes on repeat runs, 18 Oct; not scored).

One answer would have over- or under-stated those firms. Three runs give an indication, not a precise rate, which is why we report "named in X of 3 runs" rather than a position or a rank.

That first pass was also a learning run. It recorded one answer per question, not three, and some questions were pasted in an earlier wording, so we don't publish its figures. Our studies now come from a separate check on 25 Oct 2026, from a Singapore connection, with every question asked exactly as written. Those studies checked ChatGPT and Perplexity; client audits check ChatGPT and Google AI Overviews (waterproofing, renovation and interior design).

What we report: the outcome

Named in X of 3 runs

For each question and each tool: how many of the 3 runs named your firm in the answer. If a tool didn't answer, we say so, for example "named in 1 of 2 runs with an answer". For Google AI Overviews, runs where no AI Overview was shown are reported separately, for example "named in 1 of 2 runs with an AI Overview; no AI Overview shown in 1". We also note when your website or a profile about you was cited without your firm being named, and we flag negative mentions separately; they never count as a win.

Share of voice

Your firm's mentions divided by all firm mentions (yours plus every competitor named) across the whole set of questions. We list the competitors named most often, each with their own "named in X of Y runs".

Top cited sources and accuracy issues

The websites the tools cited most, which shows where it matters to be listed accurately. And anything an answer got wrong about your firm, such as an old phone number, a licence you don't hold or a wrong price. Those go to the top of the fix list.

How we read change

With 3 runs per question, a change on one question from one month to the next can be chance. We only call a change real if it holds for 2 monthly checks in a row, or if it moves by 2 or more runs on 3 or more questions. We report the numbers as they come out, including when they stay the same or go down.

The readiness score, kept separate

7 areas, 100 points

Source: weCDEF audit rubric, version 2.

The readiness score is our own checklist, not an industry standard and not a precise measurement. It covers what you have in place that search engines and AI tools can find. It isn't a prediction of how often AI tools will name you, and we never fold the outcome into it. The weights are our judgement, and we'll review them after 3 months of results.

Good, Partial, Missing, Not checked

Every checklist item gets one of four labels:

• Good: full marks for that item.
• Partial: it exists but has gaps. Part marks.
• Missing: we checked, and it isn't there. No marks.
• Not checked: we couldn't check it with our tools.

A Not checked item is left out of the score rather than counted as zero, because a gap in what we could see isn't the same as a gap in what you have. That area is scored on the items we could check, and we check the missing item again in a follow-up pass. Each area shows "scored on X of Y items checked", so you can see what's been left out. If a whole area can't be checked, the total is marked provisional and worked out over the areas we could check.

An item that doesn't apply to your work is left out in the same way and never marked down: for example, BCA's Contractors Registration System if you only work on homes. Area scores are rounded to whole points, with halves rounded up.

What we don't do

• We don't promise that any AI tool or search engine will name you, link to you or rank you first, or that you'll get more enquiries.
• We don't write, buy or swap reviews, or suggest offering anything in exchange for them.
• We don't post about you anywhere without saying who we are, and we don't create fake listings or accounts.
• Every question is asked one at a time in an ordinary browser window, signed out, and the answer is screenshotted at the time. We don't use scrapers, APIs or bulk tools on the AI tools.
• We don't sell llms.txt files or "special AI markup".

Limits we always state

• AI answers vary from run to run. Three runs give an indication, not a precise rate.
• Signed out isn't signed in. People who are signed in, with their own history and settings, may see different answers.
• Location still matters. Our checks run from a Singapore connection with Singapore named in every prompt. Homeowners in other parts of Singapore, or with different settings, may still see different answers from ours.
• AI Overviews don't always appear. Google often shows no AI Overview for a question. We record that as "No AI Overview shown", not as a miss. Other AI tools, such as Perplexity, aren't part of client audits.
• Cause and effect. A change in the answers can come from our fixes, from changes the AI companies make, or both. We report what changed; we can't always say why.
• The readiness score isn't a forecast. It describes what's in place, not what any tool will say.

FAQ

Why 3 runs and not 1? Because answers vary from one run to the next. In our first pass, the third firm named for the same question changed every time we asked it. Three runs show which names keep coming back.

Why signed out? It's a repeatable starting point with no personal history. People who are signed in may see different answers, and we say so in every report.

Why does every prompt name Singapore? Our checks run from a Singapore connection, and naming Singapore in the question as well keeps the answers about Singapore whatever location a tool guesses for us. It reduces the effect of location; it doesn't remove it.

What if Google doesn't show an AI Overview? Google doesn't show one for every search. When it doesn't, we record "No AI Overview shown" for that run. It's its own outcome, not a miss, and it isn't counted against you.

Why do your studies use Perplexity when client audits don't? Our published studies checked ChatGPT and Perplexity, on a free signed-in account, for one run per question. Client audits check ChatGPT and Google AI Overviews. We say which tools each figure comes from.

What is "share of voice"? Your firm's mentions divided by all firm mentions across the whole question set. If ChatGPT named firms 20 times and 4 of those were you, your share of voice is 20%. (That example is from our fictional sample report.)

Is the readiness score a prediction of AI mentions? No. It measures what's in place, and it's reported separately from the outcome. A better readiness score doesn't automatically mean you'll be named more often, which is why we measure both.

Do you use llms.txt or special AI schema? We don't sell either. Google says no special optimisation, files or markup are needed to appear in its AI features. If you already have an llms.txt file we note it, but it isn't scored.

Can you guarantee ChatGPT will recommend me? No, and you should be wary of anyone who does. The AI companies decide what their tools say. What we can do is fix what's within your control and measure, with the same questions each month, whether the answers change.

See it in practice

• Worked examples of the outcome data: ChatGPT and Perplexity on waterproofing in Singapore and ChatGPT and Perplexity on renovation and interior design in Singapore (25 Oct 2026, 1 run per prompt, from a Singapore connection, ChatGPT and Perplexity; a different setup from the client audits described here, which check ChatGPT and Google AI Overviews 3 times per question).
• What a report looks like: a sample report for a fictional business, and how we're testing our audit on our own sites first.
• weCDEF home