Ask an AI assistant “Which accounting firm in Lyon is best for freelancers?” or “Is this meal kit worth it?” and you will rarely get a list of ten blue links. You get a short paragraph, often with a recommendation, a caveat and a handful of cited sources.
That paragraph is a new kind of first impression. And the material it is built from is not only your website. It includes what other people write about you: reviews, forum threads, comparison articles, press coverage.
This article explains what is documented about how AI assistants use web sources, what we can reasonably infer about reviews and mentions, and how to turn that into a monitoring routine.
What is documented: assistants answer from web sources
Several assistants are explicit about drawing on the web.
Perplexity’s help center says it searches the internet in real time and that every answer includes citations linking to the original sources.
Google describes its AI features in Search (AI Overviews and AI Mode) as surfacing relevant links to help people find information. Google also states that there are no additional requirements or special optimizations needed to appear there: a page needs to be indexed and eligible to be shown in Search with a snippet.
OpenAI presented ChatGPT search as a way to get timely answers with links to relevant web sources. The practical point is the same across these products: when an assistant looks things up, it works from pages that already exist, and many of those pages are not yours.
What we can infer: reviews and mentions are part of the raw material
No assistant publishes a formula for how it weighs a Google Business Profile rating against a news article or a Reddit thread. So anything beyond “they read web sources” is observation, not fact.
Here is what tends to show up when you look at real answers across many questions:
- Questions with evaluative intent pull in third-party pages. “Best”, “reliable”, “worth it”, “alternatives to” and “reviews of” questions often produce answers that cite review pages, forums and comparison articles rather than brand homepages.
- Recurring themes travel. If many reviews mention slow delivery or excellent after-sales support, it is plausible that those themes surface in summaries, because they appear repeatedly across the sources an assistant can access.
- Absence is also a signal. A brand with few independent mentions may simply not appear in category answers, even if its own site is well optimized.
Treat these as hypotheses to test on your own market. The value is in measuring them, not assuming them.
Why this changes reputation work
Reputation used to be managed mostly for people who searched your name directly. They saw your star rating, read a few reviews and decided.
With AI assistants, reputation also affects discovery. A buyer who has never heard of you asks a category question. The assistant composes an answer from sources it finds. Whether you are named, and how you are described, depends partly on what those sources say.
That has three consequences.
1. The unit of analysis is the question, not the keyword
A ranking report tells you where a page appears for a query. It does not tell you whether an assistant recommends you, describes you neutrally, or mentions you with a warning. You need to track specific questions and read the answers.
2. Sentiment and topics matter, not only volume
A high review count with a recurring complaint can produce a less favorable description than a smaller, consistent body of positive feedback. Track what people say, not just how many said it.
3. Sources you do not control deserve attention
Comparison articles, directories, forum discussions and review platforms can carry as much weight in answers as your own pages. Knowing which ones get cited tells you where to focus outreach, corrections or content.
A practical monitoring routine
You do not need a large team to start. You need a stable list of questions and a regular rhythm.
Step 1: write the questions your buyers ask
Aim for 20 to 50 questions, grouped by intent:
| Intent | Example question |
|---|---|
| Category discovery | “What are the best project management tools for small agencies?” |
| Comparison | “Brand A vs Brand B for invoicing” |
| Trust | “Is Brand A reliable?” |
| Problem solving | “How do I reduce churn in a subscription box business?” |
| Local | “Good dentist near Bordeaux for children” |
Keep the wording close to how real people talk. Avoid stuffing your brand name into every question, or you will only measure branded answers.
Step 2: record three things for each answer
- Presence: is your brand named?
- Framing: recommended, neutral, compared unfavorably, or associated with a problem?
- Sources: which pages are cited?
Step 3: cross-check with your reviews and mentions
Read your recent reviews and mentions and list the five most frequent themes, positive and negative. Then compare them with the framing you recorded. If an assistant says “customers report slow support” and your reviews show the same theme, you have a real operational issue, not a visibility problem.
Step 4: repeat on a fixed schedule
Answers change as sources and models change. A single snapshot can mislead. Monthly is a reasonable minimum; competitive categories may justify a weekly check.
What to do with the findings
A short checklist:
- Fix the cause before the message. If a theme appears in both reviews and AI answers, address the service issue first.
- Reply to reviews consistently. Public replies show how you handle problems and add your own context next to the criticism. (See our negative reviews playbook.)
- Ask for reviews fairly. Google’s user contributed content policy prohibits incentivized reviews and selectively soliciting only positive ones. Ask all customers, without rewards.
- Strengthen the pages that get cited. If a comparison article or directory listing is cited often and contains outdated information about you, request a correction.
- Publish clear, factual pages on your own site that answer the trust questions directly: pricing, guarantees, support hours, coverage area.
None of this is new advice in isolation. What is new is that the same work now affects how you appear in answers to people who have not searched your name yet.
How SearchFirefly helps
SearchFirefly asks your tracked questions in ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode, in each market’s language, and shows for every question the brands named and the sources cited, along with your share of voice against competitors. Its reputation view brings together Google reviews, web mentions and Reddit discussions with sentiment and themes, so you can compare what people say with what assistants repeat, and prioritized recommendations point to what to fix first.