Why Do Citations Change When I Add Best, Trusted or Recommended to a Question?

How does one word change the sources?

Search-based assistants do not look up your question as typed. They rewrite it into one or more search queries, retrieve pages, and write an answer from them. Google's documentation says AI Overviews and AI Mode may use a "query fan-out" technique, issuing several related searches across subtopics and sources. The modifier usually survives the rewrite, and search engines treat "best", "trusted" and "recommended" as strong signals of what kind of page answers the query:

Models are also sensitive to wording in ways that have nothing to do with search. Research on prompt formatting found performance differences of up to 76 accuracy points on one open model from changes that did not alter the meaning. A modifier is a bigger change than formatting, so large shifts in citations are unsurprising.

Is it the modifier or just noise?

Both, usually. Citations vary between identical runs, so a single comparison proves nothing. Measure it properly:

1. Pick pairs: the plain question and each modified version, such as "project management software for agencies" and "best project management software for agencies".
2. Run each version at least five times in the same assistant, same settings, same day.
3. Record the cited domains for every run.
4. Calculate overlap within each version: for each pair of runs, divide the domains cited in both by the domains cited in either. This is the Jaccard overlap. Average it across pairs. This is your noise baseline.
5. Calculate overlap across versions the same way, pairing runs of the plain question with runs of the modified one.
6. Compare. If the overlap across versions is close to the overlap within versions, the modifier is not doing much. If it is much lower, the modifier is changing the sources.

As an example, if two runs cite five domains each and share three, the overlap is 3 ÷ 7, about 0.43. Suppose runs of the plain question average 0.48 overlap with each other, and runs of the "best" version 0.45. Plain-against-"best" pairs average 0.18. Most of the difference then comes from the word, not from noise. How to size the noise itself is covered in how to measure the noise floor for ChatGPT citation tracking.

What should you look at besides overlap?

Classify each cited source by type: your own pages, competitor pages, roundups and listicles, review platforms, news, forums and community threads, professional bodies, and other. Then compare the mix for each wording. The mix tells you what kind of evidence the modifier rewards, which is more useful than which exact domains appeared. A study of cited sources for branded questions, drawn from 23,387 unique sources, found that reviews and social proof made up 57% of citations. That group covered reviews, listicles, forums, social posts and case studies, against 12% for product and commercial pages. Modifiers such as "best" and "trusted" are likely to push the mix further toward that kind of source; the source-type comparison below shows whether they do for your questions. How to classify sources consistently is covered in how to classify owned, earned, neutral and competitor citations.

What should you do for each modifier?

For "best": your presence depends mostly on third-party roundups and comparisons. An observational study of about 200,000 AI responses found that appearing in frequently cited third-party listicles was strongly associated with being mentioned. Ranking first added 16.5 percentage points to the mention rate in B2B SaaS, and the benefit levelled off after a few placements. Find the roundups cited for your "best" questions and pursue inclusion on merit. On your own site, publish honest comparison pages. Lists of "best" products that always rank the publisher first have been reported to lose visibility.

For "trusted": publish the evidence of trust, in text: certifications with their issuers and verification links, security and compliance pages, years in business, customer numbers you can support, and regulator registrations. Build review platform profiles with a steady flow of genuine reviews. Earned news coverage also counts here.

For "recommended": assistants look for endorsements by others. Professional body listings, expert recommendations, case studies with named customers, and genuine discussion in communities where your buyers ask each other for advice all matter. You cannot write these yourself, but you can make it easy for customers and experts to recommend you. Give them clear product names, shareable case studies and a reason to talk about you.

The difference between these outcomes and simply being cited is covered in the difference between being cited by AI and being recommended by AI.

Which wordings should you track?

Track the ones your buyers use. Look at your site search, sales call notes and support tickets for how customers phrase their questions. Include at least the plain version and "best" for each core question, and add "trusted" or "recommended" where buyers care about risk, such as in finance, health, security or professional services. Keep the set fixed over time so changes reflect the market, not your wording. How visibility changes between comparison, list and category questions is covered in why your AI visibility changes between versus, listicle and category prompts.

What might an illustrative case look like?

Consider an illustrative cybersecurity training company that appeared in most answers to "security awareness training for small businesses" but rarely when "best" or "trusted" was added. It ran each version five times in three assistants and found overlap across versions far below overlap within them. The plain question cited vendor pages, including its own. The "best" version cited four roundup articles, none of which listed it. The "trusted" version cited review platforms, where it had few reviews, and certification pages of competitors. It pitched the roundup authors with a product trial, and published its certifications with verification links. It also began asking customers for reviews after onboarding. Four months later, it appeared in half of the "best" answers, cited through two roundups, and in a third of "trusted" answers.

What do people ask about modifiers and citations?

Should we add "best" and "trusted" to our own page titles?

Only where accurate and natural. Self-promotional "best" lists do get cited (Ahrefs found them the most prominent page type in ChatGPT's sources), but they have been reported to lose Google visibility, and an answer can cite your list yet recommend a competitor. Earning a place in third-party roundups is the safer route.

Do all assistants react the same way?

No. Each rewrites and retrieves differently, so measure each assistant separately.

How many wordings should we test?

Two or three per core question is usually enough to see the pattern without making the set unmanageable.