Original insights by Tomek Rudzki
30-second rundown
Key learnings:
- ChatGPT rewrites recommendation questions: In Rudzki’s 5-million-query analysis, hidden searches often added “best,” comparisons, reviews, brands, and dates before the answer was assembled.
- One broad page cannot cover the whole search chain: Build evidence for the repeated angles the model investigates, including use cases, comparisons, credible reviews, and current information.
- Different answer engines investigate differently: At the time, ChatGPT averaged 2.1 extra searches per prompt, Perplexity 1.4, and Grok 6.8, so visibility tactics must reflect the model being measured.
This week: Choose five important buyer questions and record the hidden searches produced for each using FanoutFox. Group the repeated additions into comparisons, use cases, reviews, brands, and freshness, then improve the page that misses the most common angle. Finish with a short map showing which page or third-party source supplies evidence for every repeated search.
Coming 3 months: Set up an AI workflow to collect fan-outs for a stable set of buyer questions every month, group recurring themes, and flag missing evidence or outdated pages. Assign one owner to approve the highest-value fixes and track whether your pages and trusted third-party mentions begin appearing across more of the hidden searches.
Tomek Rudzki watched five million hidden searches crawl out from behind ordinary questions, and the result looks less like a search box than an interrogation room with bad wiring. Ask ChatGPT for a recommendation and it may quietly summon comparisons, reviews, brands, current dates, and a jury of websites before it gives you an answer.
These hidden searches are called query fan-outs. One user question branches into several narrower searches so an answer engine can gather evidence, compare options, and decide what deserves to appear.
The original question is only the opening witness. If your content answers those exact words and nothing else, your brand may disappear during the private questioning that follows.
1. The Question Gets Rewritten Behind Your Back
Rudzki analyzed five million fan-outs collected from April 1 through April 21, 2026 across ChatGPT, Perplexity, and Grok. Most came from ChatGPT, so his headline findings describe how that system expands a question before assembling an answer.
He reports that ChatGPT uses reciprocal rank fusion, a method that combines results from several searches. In plain English, a source that performs well across multiple related searches has a better chance than one that appears for only a single narrow angle.
A page about family electric cars may therefore need credible information on range, safety, price, performance, and reviews. Answer only “best family electric car” and the hidden search party may ride straight past your fence.
Coverage must survive several related searches. This does not mean stuffing every phrase into one swollen page. It means building a connected body of useful pages and outside evidence that answers the comparisons and verification questions the model actually asks.
2. “Best” and “Reviews” Enter the Room Uninvited
The most common word ChatGPT added was “best,” followed by other commercial terms such as “top,” “comparison,” “reviews,” “tools,” “software,” and “features.” For advice-style questions, “best” appeared in 24.3% of fan-outs. That means nearly one in four hidden searches in this group turned ordinary advice into a contest.
Reviews were the third most frequently injected word, even when the user never asked for them. The machine wants corroboration, preferably from people and platforms outside the seller’s own compound. Your product page can praise itself until the coyotes file a noise complaint, but an answer engine may still go hunting for an independent verdict.
Freshness joins the raid too. ChatGPT added the current year in 5.44% of prompts, while Grok did it more often. The practical lesson is not to repaint every page on the site, but to update the pages already influencing AI answers and keep their evidence visibly current.
Recommendation visibility depends on outside proof. Build honest comparisons around specific needs, summarize credible review evidence, watch smaller review sites as well as famous ones, and correct real product problems instead of trying to perfume the corpse.
3. Each Answer Engine Runs a Different Investigation
In April 2026, the average prompt produced 1.4 fan-outs in Perplexity, 2.1 in ChatGPT, and 6.8 in Grok. Those numbers do not make one system automatically better. They show that each system breaks down questions with a different appetite and method.
Perplexity usually simplified the original wording. ChatGPT stayed close to the intent but added brands, comparisons, “best,” and reviews. Grok behaved like a caffeinated research assistant, widening and narrowing the search with dates, brand-versus-brand queries, and instructions to search particular websites.
Grok used a search instruction aimed at a specific site in 18.3% of chats. Reddit appeared in 10.5% of all chats, and nine out of ten of those appearances were deliberate searches of Reddit rather than casual mentions. Rudzki also found Grok independently targeting sources such as Wirecutter and Consumer Reports for product evaluations.
There is no universal fan-out playbook. Measure the answer engines your customers actually use, inspect the sources and angles each one adds, and build evidence for that real search behavior instead of worshipping one blended dashboard number.
The visible prompt is the front door. The real decision may happen in six back rooms where comparison pages, review sites, current dates, community opinions, and trusted publishers are all asked whether your brand belongs on the list.
Map those rooms. Put useful evidence in each one. Keep it current and let independent sources carry the parts of the case your own website cannot credibly argue.
Answer the investigation, not just the question. Otherwise ChatGPT may hear the customer say your category, conduct a complete trial in secret, and return with somebody else’s name.
