Our work

Veesie: measuring whether AI mentions your brand, every week, with the noise taken out

Veesie is a SaaS brand of our own that we built from scratch. It asks ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews the questions your customers ask, with the model's search function switched on, and counts who appears in the answer. Every week, in Dutch, French and English, and set up so that one chance answer does not decide your figure.

Ask an AI assistant the same thing twice and you get two answers. Measure that without repetition and you are measuring chance.

Challenge

More and more customers ask an AI assistant for a recommendation instead of scrolling through ten blue links. So a business wants to know whether it appears in those answers, next to whom, and whether that is improving. The catch: an AI answer differs per attempt, per model and per language, which makes a snapshot worth very little. The question was whether you can turn that into a figure that means something, and do it again every week.

Approach

Real questions, asked live with the model's search function on, across five sources. Repetition where the models vary most, a fixed basket of three sources for the score so your line does not jump, and everything that needs no judgement in code. Built with Claude Code, and guarded by a test suite that grew with every feature: 521 test files at the time of this case.

5

AI sources per question: ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews

4 points

between the highest and the lowest of five identical measurements on one day (25 August 2026, 12 questions): 85, 87, 84, 88 and 85

0

missed mentions when reading 655 answers (test of 24/08/2026)

521

test files in the repo, with the prices, plan limits and chatbot knowledge as guarded facts

Real questions, with search switched on

Every question goes live to the source, with search switched on: gpt-5-search-api for ChatGPT, sonar for Perplexity, Claude Haiku 5.5 and Gemini 2.5 Flash-Lite with web search, and Google AI Overviews straight from the real Google results through DataForSEO. When the model cites sources, Veesie stores them with the answer, as a clickable link. So you see not only whether you are mentioned, but also which pages the AI read to get there.

For every answer, Veesie records whether the brand is mentioned, in which position, in what context and with what sentiment. The extraction was tested on 655 answers over seven days: zero missed mentions that were literally in the text, zero mentions without grounds.

One measurement is noise, five show the pattern

On 25 August 2026, Veesie ran five identical measurements on clickforest.com: 12 questions, five sources, within one hour, with the models of the time (Claude Haiku 4.5; since 8 October 2026 Veesie measures with Haiku 5.5). The score stayed stable: 85, 87, 84, 88 and 85. Underneath, plenty changed. In 14 of the 60 combinations of question and source, the answer came out differently from one run to the next, and 10 of those 14 came from Claude and Gemini.

0 50 100 85 87 84 88 85
Five identical measurements on clickforest.com, 25 August 2026, 12 questions.

That test is about how stable the figure is for the same questions, not about the level. The dashboard in the gallery shows a new measurement from 10 October 2026 on 17 different questions, which Veesie derived from the site, and there clickforest.com scores 70. A different question set gives a different level.

Two design choices followed from this. Claude and Gemini get every question up to three times, and Veesie keeps the majority. And weekly became the maximum on every plan: measuring more often barely changes the figure, what does change is your site. For that there is an extra measurement on request, after a change.

A score that does not jump

The GEO Score is an index of our own from 0 to 100, which is why the formula is public (veesie.com/features): 50% mention rate, 25% sentiment, 15% source coverage and 10% position in the answer. On every plan the score is calculated over the same three sources: ChatGPT, Perplexity and Google AI Overviews. Claude and Gemini are measured and shown from Plus, but do not count. Anyone who changes plan therefore does not see their line jump because of a different basket.

From measurement to advice

Measuring is half the job. When you add a website, Veesie reads the site itself and suggests questions of three kinds: broad, per service and local. From Plus, the AI-Readiness Audit checks the site on 20 checks, ten for SEO and ten for GEO, and puts the most important improvements at the top. From Pro, Veesie writes GEO Recommendations based on the last five measurements, puts them in an action plan, and the GA4 connection shows in your own visitor figures how much traffic AI brings in. That connection reads Google Analytics with the scope analytics.readonly, verified by Google on 24 August 2026.

How Veesie compares with other tools is covered in our comparison of GEO monitoring tools.

Built with AI, guarded by tests

Veesie was built with Claude Code, from the first commit on 25 April 2026 to the public launch in October: 1,927 commits and 100 database migrations. What makes it reliable is the test suite underneath. Besides the measurement logic, it also checks the facts visitors read. Prices and plan limits come from one configuration file, and tests fail when a text or the chatbot promises a feature the plan does not have, or names a number that is no longer correct. A chatbot that does not know a feature does not say "I don't know". It answers just as confidently from what it does have. So its knowledge gets tested too.

A GEO tool has to be cited correctly itself

A tool that measures AI visibility is itself described by AI. On 9 October 2026, an AI assistant showed Veesie in a comparison with the wrong logo and quoted the yearly price as the monthly price. The cause was probably on our own site: a schema markup that described two brands as one entity, and a pricing page that showed the yearly price large without saying so. The same day there were citable sentences per plan, FAQs on exactly those misunderstandings, a corrected schema and an IndexNow ping for all 69 pages. Tests make sure it does not come back.

What it costs to run

Veesie runs the cheapest model per family, with search switched on. With web search, the search layer shapes the answer more than the size of the model, and that makes a weekly measurement possible from €29 a month. The cost per measurement is not estimated but calculated from real measurements and the rates per model, and checked again at every model change. The measurements themselves run through Inngest, with a separate step per combination of question and source, so a measurement of 85 combinations does not run into the time limit of a single Worker.

What we do not claim

  • No customer results. Veesie has been open to the public since October 2026. Figures on what customers achieved with it will come once they exist.
  • Not a mirror of the consumer app. Veesie runs the cheapest model per family, with search switched on. Which brand a model names first can differ from what someone sees in the app today. As a relative measure, week after week, the measurement is reliable.
  • One stability test, on one site. The series 85, 87, 84, 88 and 85 comes from clickforest.com, a brand that gets mentioned often. For a brand that hardly appears anywhere, the figure jumps relatively more: on veesie.com itself, the same kind of measurement went from 0 to 22.
  • No industry standard. The GEO Score is an index of our own. That is why the formula is on the site, so you know what a rise means.
  • Veesie itself is hardly mentioned yet. A new brand still has to build that visibility, and Veesie measures that for itself too, every week.
  • The dashboard is in Dutch for now. Veesie measures questions in Dutch, French and English, but the dashboard itself is currently only available in Dutch.

Five sources, every week

ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews, queried live with search switched on, in Dutch, French and English.

Repetition against noise

Claude and Gemini get every question up to three times. Five identical measurements on one day gave 85, 87, 84, 88 and 85.

Facts that get tested

Prices, plan limits and what the chatbot says come from one source and are guarded by the test suite.

Tools used

Claude Code (building with AI)
Next.js 15 and React 19, through OpenNext on Cloudflare Workers
Supabase (Postgres and Auth) in Frankfurt
Inngest for the measurements, with one step per question and source
OpenAI, Anthropic, Google, Perplexity and DataForSEO as AI sources
Stripe for the subscriptions, Resend for the emails
Cloudflare Turnstile and Sentry

Good to know before you start a free trial: the Veesie dashboard is currently available in Dutch only.

Your project as the next case?

We work with Belgian SMEs that want to grow.

Through Shopify, performance marketing and AI.

Client reviews

What clients experience

Why entrepreneurs love working with ClickForest.

Excellent on Google 5.0 ★★★★★ based on 25 reviews