11 min read
TreeHouse traced AI referrals to paid subscription trials and reports that AI search now supplies about 25% of its organic conversions. The case study shows a channel worth measuring, and a lift claim would need a comparison it leaves out.
TL;DR Tap to expand #
- TreeHouse, the Dutch company behind Pararius and Huurwoningen.nl, replaced manual ChatGPT and Perplexity checks with automated prompt monitoring, crawler logs and referral tracking that ends at a paid subscription trial.
- Promptwatch, TreeHouse’s monitoring vendor, reports that AI search supplies roughly 25% of TreeHouse’s organic conversions and that ChatGPT accounts for 99% of the AI-referred ones.
- The 25% is a share of attributed conversions. The case study publishes no absolute counts, baseline or control, so it describes the channel mix and leaves lift unmeasured.
- You can copy the move by tying AI referrals to one paid event before you optimise for mentions or citations.
Experiment
25% of organic conversions attributed to AI Search
Connect AI referrals to a commercial event before optimizing citations, and treat channel share as attribution rather than incremental lift.
Picture a marketing team that can watch ChatGPT recommend its brand all day and still cannot say whether one of those recommendations ever paid the bills. TreeHouse lived there. The Rotterdam company runs Pararius and Huurwoningen.nl, two rental platforms where tenants meet landlords, and its marketers were putting the same three prompts to ChatGPT and Perplexity hundreds of times a day, then pasting the answers into spreadsheets across four platforms. The routine surfaced visibility and hid revenue.
Marketers have a name for the missing piece. It is AI search attribution, the work of connecting a prompt, a cited source and a referral visit to a conversion your finance team recognises. If you run a company rather than a channel, the plain version goes like this. Someone asks an AI assistant where to rent a flat, your brand gets named, and you want to know whether that person later paid you.
A case study that Promptwatch published on 9 October 2026 says TreeHouse built that connection. Promptwatch is the software vendor behind TreeHouse’s monitoring, so it is both the source and a party to the result. The case study credits AI search with about 25% of TreeHouse’s organic conversions and says ChatGPT supplies 99% of the conversions that arrive through AI referrals. The conversion in question is a paid subscription trial.
The gap sat between the citation and the trial #
TreeHouse develops and operates platforms for the rental and purchase property market. Its brands include Pararius and Huurwoningen.nl, which connect prospective tenants with landlords and property professionals. TreeHouse says it employs more than 80 people, and Pararius describes a wider group that spans four rental websites and other real estate products.
That portfolio handed the marketing team a measurement puzzle. SEO (search engine optimisation) brought meaningful demand, yet conventional rank trackers could not show which brands a large language model, or LLM, recommended. An LLM is the engine behind assistants such as ChatGPT. The trackers also could not show which sources the model cited or which prompt came before a visit. A brand mention inside an answer can look like a win and never produce a single trial.
So the team started by hand. According to Promptwatch, they asked ChatGPT and Perplexity a small set of prompts again and again and moved the answers into spreadsheets. That gave an early signal, but it would not scale across four platforms and could not explain why visibility moved. The documented constraint sat in measurement, in the chain between an AI answer and the paid event at the end of the funnel.
TreeHouse built the measurement chain before chasing mentions #
The team opened with branded prompts to learn what the models said about each platform. It then widened to organic and competitor prompts, using Promptwatch’s Google Search Console integration (Google’s free report on how a site performs in search) to help choose what to monitor. Hundreds of prompts a day replaced three in a spreadsheet.
Crawler logs came next. They show which AI bots reach your site and where broken pages, access rules or conflicting information limit what those bots can read. TreeHouse allowed the relevant crawlers, fixed 404 errors (pages that return a not-found message), removed thin or outdated pages and corrected facts that disagreed across pages. It also added noindex tags where they fit, which tell search engines to leave a page out of their index.
This is the point where LLM referral tracking became a routine rather than a dashboard. Prompt monitoring flagged visibility changes, crawler logs exposed access problems, web analytics followed each referral and the paid trial closed the loop. TreeHouse later connected the Promptwatch MCP to Claude, so the team could investigate traffic spikes and receive a weekly report on what moved and what to check next. MCP stands for Model Context Protocol, a standard that lets an AI assistant read data from another tool.
The case study states no formal hypothesis, so this one is my reading of the sequence. Better access, cleaner facts and more original material worth citing should raise useful AI visibility, and the referral tracking shows whether that visibility reaches paid trials.
Two sister brands pointed to the pages AI cites #
The portfolio supplied a handy comparison. Promptwatch says TreeHouse set Huurwoningen.nl beside Pararius and saw that the brand publishing more market reports and proprietary rental data earned more citations from AI systems. TreeHouse took that as a reason to prioritise original data over rewrites of existing pages.
Treat it as a lead and keep it out of the proof pile. The two brands differ in history, authority, audience, content and distribution, so the comparison cannot isolate proprietary data as the cause. What it gave TreeHouse was an editorial choice anyone could audit, which is how generative engine optimization starts to look like a working practice. That term covers shaping content so AI assistants cite it. You find the pages and datasets the models cite, then check whether those referrals go on to convert.
Geography matters too. TreeHouse runs rental platforms in several national markets, and its public material says the group has expanded beyond the Netherlands into the United Kingdom and France. Language, local rental rules and country-specific data make one global prompt set a weak stand-in for each market. A comparable operator would monitor and track conversions separately per country and avoid treating an English-language result as the whole channel.
AI search now supplies about a quarter of organic conversions #
Promptwatch reports that AI search accounts for approximately 25% of TreeHouse’s organic conversions and that ChatGPT supplies 99% of the conversions that come from AI referrals. Because the endpoint is a paid subscription trial, the number carries more weight than an impression count or a citation tally.
The case study leaves several figures out. These include how many trials sit behind the percentage, what share of organic traffic comes from AI, which platforms and countries contributed, and the exact measurement window. Dave Vollebregt, TreeHouse’s marketing manager, says traffic and conversions rose significantly over the last few months, though the starting level and the pace of change stay unpublished.
Founders, this is the point to slow down. A channel can reach 25% of organic conversions because it grew fast, because conventional search shrank, because the attribution rules changed, or because AI visitors arrive with unusually high intent. Marketers will recognise the problem as share versus incrementality. Share tells you where conversions come from. Incrementality tells you how many conversions would have vanished without the work. The public evidence on AI search conversions supports the first and says nothing yet about the second.
Read the 25% as share, then decide what to test #
The TreeHouse case has no untreated brand, randomised page set or stable before-and-after comparison. It bundles monitoring, crawler access, 404 repairs, content pruning, noindex changes, factual corrections and more proprietary reporting into one programme. Any growth could trace back to one of those changes, to wider consumer adoption of AI search, or to a mix.
Even so, the finding earns its place as attribution. TreeHouse learned that ChatGPT referral traffic reached a business event in enough volume to win a place on the technical roadmap, and the team swapped hours of manual checking and reporting for a repeatable data workflow. The same reading question came up when a small ChatGPT ad test drove leads, where the signal also needed a careful look.
Attribution can change a decision without proving incrementality. RS Group met a related boundary when it fed reconstructed journeys into paid-search bidding, and in both stories the measurement layer became part of the growth system. The causal claim still waits on a comparison that the public TreeHouse case leaves out.
A copyable audit starts with the paid event #
You can copy the operating move without building a model. Pick one conversion that finance or product analytics records, such as a paid trial, a qualified lead or a completed purchase. Keep the referring URL and landing page, then tag visits from known AI surfaces separately from traditional organic search.
Next, connect that event to a deliberately small prompt set covering branded prompts, category prompts and competitor comparisons, split by language and market. Record the answer, the cited source, the model and the timestamp. The aim is to learn which observed citations come before commercial visits and which never do, with a bigger visibility score as a side effect at best.
Run a technical audit beside the prompt monitor. Check whether relevant crawlers can reach the cited and uncited pages, whether old URLs return errors, whether similar pages contradict each other and whether outdated material can still be indexed. TreeHouse’s cross-brand observation suggests proprietary data deserves attention, so treat it as a content type to test and leave the universal ranking rule out of it.
Then keep one comparison alive. You could phase technical fixes by site section, hold back a new market report for a matched group of topics, or use a time-based switchback, which switches the change on and off in set periods, where the intervention allows it. Agree the primary outcome and the minimum observation window before anything changes. Referral conversions, assisted trials and conversion rate all help, with overall organic demand and duplicate attribution as guardrails. A holdout, meaning a group that never sees the change, is the cleanest comparison, though Cann’s 19 day of its top ad channel shows how messy one can get.
Keeping the chain honest #
LLM answers shift between runs, logged-in states, locations and model versions. Referral URLs also miss journeys where someone reads an answer and types the brand name into a browser later. A tidy dashboard can therefore undercount exposure and still overclaim causation.
TreeHouse’s order of operations is worth borrowing. It set a baseline, automated the observing, repaired access and content problems, followed referrals to a paid event and used the result to reset internal priorities. Which of those steps produced how much extra demand remains open, and answering it takes a counterfactual, meaning a defensible picture of what would have happened without the work. Citations show where a model found its evidence and referrals show which visits arrived, and only the counterfactual shows what the work added.
Sources #
- [Promptwatch case study on TreeHouse’s AI Search workflow](https://promptwatch.com/case-studies/how-treehouse-grew-llm-traffic-and-conversions-with-promptwatch) , published 9 October 2026
- [TreeHouse company website](https://treehouse.nl/) and[company profile](https://treehouse.nl/over-ons)
- [Pararius company profile](https://www.pararius.com/info/about-pararius)
- [Dave Vollebregt of TreeHouse on LinkedIn](https://nl.linkedin.com/in/davevollebregt)
FAQ #
Compare the evidence and operating lessons from more AI experiments in our Growth Signal Index and the Growth Signal archive, or browse benchmark tooling in the Industry Contents Lab.