Someone types into ChatGPT: "What's the best window replacement company in Bellingham?"
Thirty seconds later, three businesses are named. One gets a glowing recommendation. One is mentioned as an alternative. One is conspicuously absent — despite having 400 five-star reviews on Google.
This isn't random. AI engines follow a systematic process to decide which businesses get recommended, and it's fundamentally different from how traditional search engines rank websites. There is no page-one equivalent. There's no algorithm you can game with backlinks. And your Google ranking has surprisingly little to do with it.
We analyzed over 400 AI-generated responses across four major engines — ChatGPT, Google Gemini, Perplexity, and Claude — to reverse-engineer how each one selects and ranks business recommendations. What we found is a five-stage signal chain, a small set of "gatekeeper" sources that control visibility, and significant disagreement between engines about who the best businesses actually are.
The Five-Stage Signal Chain
Every AI recommendation passes through five stages before it reaches the person who asked the question. A breakdown at any single stage means your business doesn't get mentioned — regardless of how strong you are at the other four.
Query Intent Classification
Before the engine does anything else, it decides how to answer the question. This is the most underappreciated stage because it's invisible to the end user, and it determines whether your web presence matters at all.
- Web search: The engine queries the live web, retrieves sources, and synthesizes an answer. Your website, your reviews, and your presence on third-party sites all matter.
- Training data only: The engine answers from its pre-existing knowledge without searching. Your current web presence is completely irrelevant — all that matters is what was in the training data.
- Hybrid: Some combination of web search and training data.
The split varies dramatically by engine. Claude only searches the web for roughly 60% of queries — the other 40% come entirely from training data. Google Gemini uses "Search Grounding," which queries Google Search for nearly every factual question. This means the same business can be visible on one engine and completely invisible on another, not because of content quality, but because one engine searched the web and the other didn't bother.
Source Selection
When the engine does search the web, it retrieves between 5 and 10 source articles to inform its response. These sources become the engine's window into your industry — and this window is far narrower than most people realize.
- The top 5 most-cited source domains account for approximately 40% of all citations across our dataset.
- Source overlap between engines is remarkably low — only 11% domain overlap between ChatGPT and Perplexity.
- Each engine has strong preferences: Perplexity leans heavily on third-party editorial sites, while Gemini favors traditional SEO winners from Google's index.
This is where "gatekeeper" sources become critical. If the two or three editorial sites that an engine trusts for your industry don't mention your business, you likely won't appear in the answer — no matter how good your own website is.
Brand Extraction
Once the engine has its source articles, it needs to extract specific business names to include in the answer. This is where many businesses silently fail. Not every brand mentioned in a source article survives into the AI's response.
- We observed a 55–65% survival rate from source mention to AI answer.
- Brands with specific, quotable facts survive at much higher rates. Things like "$7/user/month," "4.7 out of 5 on G2," or "best for teams under 20" give the AI something concrete to latch onto.
- Vague mentions get filtered. "They also offer project management solutions" rarely makes it into the final answer.
The extraction filter rewards specificity. Generic descriptions of what your business does are functionally invisible to this stage of the pipeline.
Ranking and Framing
Surviving extraction isn't enough. The engine now assigns each brand a position in the response and frames it with language that shapes perception.
- Typically, an engine gives 1 primary recommendation per 5–7 mentioned brands. The rest are "alternatives" or "also worth considering."
- Position matters: the first-listed brand carries approximately 3x the influence of position 3 or later. Being recommended and being the first recommendation are meaningfully different outcomes.
- The framing language determines which queries you win. "Best for small teams" and "best for enterprise" are completely different slots, and the engine decides which one you get.
User Response
The final answer is delivered to the user. And here's the critical shift from traditional search: there is no click-through to optimize for.
- 93% of AI-assisted searches end without a click to an external website (Bain, 2025). The user gets their answer and acts on it directly.
- Being in the answer is the conversion event. There's no "ranking higher in results" — you're either recommended or you're not.
- The framing the engine gives your brand is the first impression. You don't get to control it with your own landing page copy.
The Gatekeeper Effect
The most surprising finding in our research is how few sources control the entire AI recommendation pipeline for any given industry. We call this the gatekeeper effect: a small number of editorial websites act as intermediaries between your business and AI engines, and if you're not featured on those sites, you are functionally invisible.
In the project management software category, the numbers are stark. Zapier was cited in 14 out of 30 Perplexity answers about PM tools, making it the source behind nearly half of all recommendations in that space. The top five cited sources accounted for roughly 40% of all citations.
The Zapier Paradox
Here's the twist: Zapier is the number one cited source in PM software, but it is never recommended as a product. Zapier is a citation powerhouse but a product ghost. This proves something important: being a cited source and being a recommended product are entirely separate signals in the AI pipeline.
What happens is this: when Perplexity or ChatGPT retrieves Zapier's "Best Project Management Tools" listicle, the brands featured favorably in that article inherit Zapier's trust. The AI doesn't recommend Zapier — it recommends the products Zapier says are good.
Key finding: Your goal isn't to become a source that AI engines cite. It's to be featured favorably in the sources that AI engines already trust for your category.
Content Types That Drive AI Citations
Not all content formats carry equal weight in the AI pipeline. Across our dataset, the breakdown was clear:
| Content Type | Share of Citations |
|---|---|
| "Best X" listicles | ~74% |
| Head-to-head comparisons | ~15% |
| Review aggregator pages | ~5% |
| Vendor comparison pages | ~4% |
| Forums (Reddit, Quora) | ~2% |
Nearly three-quarters of all AI citations trace back to listicle-style content. If someone publishes a "Best Window Companies in Seattle" article and it ranks well enough for AI engines to retrieve it, the businesses listed in that article are the ones that get recommended. The businesses not listed don't exist in that engine's world.
Engine-by-Engine Behavioral Profiles
One of the most persistent misconceptions about AI search is that it's a single channel. It's not. Each engine has a distinct personality — different source preferences, different levels of opinionatedness, different stability characteristics. Optimizing for one does not automatically help you on another.
Google Gemini (AI Overviews)
- Uses "Search Grounding" which queries Google Search directly, making it the closest proxy for Google AI Overviews.
- Tends to present options neutrally rather than picking favorites — you're more likely to see "here are some options" than "the best choice is."
- Source selection correlates most strongly with traditional SEO rankings. If you rank well on Google Search, you're more likely to appear in Gemini answers.
- Best starting point for businesses already investing in SEO — there's meaningful carryover.
Perplexity
- The most opinionated engine in our testing — will confidently pick a favorite and explain why.
- Overwhelmingly favors third-party editorial sources. In PM software, Zapier content powered roughly 70% of Perplexity's recommendations.
- Community content from Reddit and forums has a notable influence that isn't present in other engines.
- The most different from traditional Google Search results. High Google rankings do not predict Perplexity visibility.
- Least stable — the same query can yield noticeably different recommendations across sessions.
ChatGPT
- The most generous engine — tends to list the most brands per response, giving more businesses a chance to appear.
- Draws from a mix of editorial sites, Reddit discussions, and review platforms. No single source type dominates as strongly as Zapier does on Perplexity.
- Highly consistent: the same query returns substantially similar results ~90% of the time, making it the most predictable engine to audit.
- Appends
?utm_source=chatgpt.comto outbound links, making it the only engine where you can directly track AI-referred traffic in your analytics. - Only 11% source domain overlap with Perplexity — they are effectively different markets pulling from different information pools.
Claude
- The most stable engine in testing — near-identical answers for the same query ~95% of the time.
- Only searches the web for approximately 60% of queries. The remaining 40% are answered entirely from training data, meaning your current web presence has zero impact on those answers.
- For local business queries, appears to use a tri-modal system: web search results, training data knowledge, and local information databases.
- The training-data dependency creates a unique challenge: even if you improve your web presence today, Claude may not reflect those changes until its next training update.
The bottom line: These four engines are four separate markets with different source preferences, different stability profiles, and only ~11% source overlap. A strategy that works for Gemini may be invisible to Perplexity.
The Review Volume Myth
One of the most common assumptions businesses make is that review volume drives AI recommendations. More reviews, more visibility. It's intuitive — and it's wrong.
We tracked review counts on G2 (one of the major business software review platforms) against AI recommendation rates across all four engines for project management tools. The results break the assumption cleanly:
| Brand | G2 Reviews | AI Visibility |
|---|---|---|
| Smartsheet | 21,442 | 33% |
| Monday.com | 15,073 | 50% |
| Asana | 13,321 | 77% |
| Notion | 10,844 | 20% |
Smartsheet has more G2 reviews than any other PM tool — and the lowest AI visibility rate in its cohort at 33%. Notion, despite nearly 11,000 reviews, is recommended in only 20% of AI responses. Meanwhile, Asana achieves 77% visibility with fewer reviews than either of them.
The correlation between review volume and AI recommendation rate isn't just weak — it's slightly negative in this dataset. More reviews doesn't mean more AI visibility.
What does correlate? Presence in the editorial articles that AI engines actually cite. Asana appears consistently in the "Best PM Tools" listicles from Zapier, PCMag, and Forbes Advisor — the gatekeeper sources for this category. Smartsheet, despite its massive review count, is frequently absent from those same articles. The reviews exist on G2, but the AI engines aren't looking at G2 for their recommendations — they're looking at the editorial listicles that may or may not mention G2 data.
Key finding: AI engines don't count your reviews. They read the editorial articles that may reference your reviews. The intermediary matters more than the raw signal.
What This Means for Your Business
The five-stage signal chain creates a clear set of priorities for any business that wants to be visible in AI-generated recommendations:
1. Identify your category's gatekeeper sources. Every industry has a small set of editorial sites that AI engines trust. For SaaS products, it might be Zapier, G2 editorial, and PCMag. For local services, it might be Yelp, Thumbtack, and local media. These are the sites you need to be featured on — your own website alone isn't enough.
2. Get specific and quotable. The brand extraction stage rewards concrete facts, not generic descriptions. "Serving the Pacific Northwest since 1987" survives extraction. "We provide quality service" does not. Price points, ratings, years in business, specific service areas, industry awards — these are the details that make it through the filter.
3. Think in four markets, not one. ChatGPT, Gemini, Perplexity, and Claude draw from different source pools with only ~11% overlap. A strategy that makes you visible on Gemini (SEO-heavy) may leave you invisible on Perplexity (editorial-heavy). You need to know where you're strong and where you're missing on each engine independently.
4. Own your framing. AI engines don't just decide whether to mention you — they decide how to describe you. "Best for enterprise teams" and "most affordable option" are different positioning slots. The framing in your source articles shapes the framing in AI answers. Make sure the gatekeeper sources describe you the way you want to be described.
5. Stop chasing review volume as an AI strategy. Reviews matter for trust signals and conversion, but they don't drive AI recommendations. The editorial intermediaries between your reviews and the AI engine are what matter. A glowing mention in one Zapier listicle does more for your AI visibility than 5,000 additional G2 reviews.
Find out where AI engines are recommending your competitors instead of you
A BeCited audit maps exactly which sources AI engines cite for your category, which ones mention your competitors, and where you're missing from the conversation.
Get Your Free Check