AI Visibility · Mechanism
Same businesses, same category, four different answers. Here is what each engine is actually weighing, and why consistency moves all of them.
By Izzy Gregorio · Updated August 2026 · 9 min read
In short
AI chatbots recommend brands using two inputs: what the model absorbed during training, and what it retrieves live from the web at the moment of the question. Each engine weights those inputs differently, which is why ChatGPT, Perplexity, and Gemini often name different brands for the same question. Consistency across sources is what moves all of them.
Key takeaways
ChatGPT names 1.2 percent of local business locations. Gemini names 11 percent. Same businesses, four very different thresholds.
OpenAI runs three separate crawlers you can control independently. Blocking the wrong one silently removes you from citation eligibility.
Roughly four out of five sources cited about a brand are pages the brand does not own.
Answer engines are corroboration machines. You persuade them by being confirmed elsewhere, not by asserting harder on your own site.
The mechanism
Model priors and live retrieval. Everything else is detail.
Model priors are what the system absorbed during training. If your brand appeared consistently across the web in the years before a model's training cutoff, described the same way, associated with the same category, the model carries a stable internal association. This is why long-established brands surface in answers even when their websites are technically mediocre. They were described consistently by a lot of sources for a long time.
Live retrieval is the search the system runs at question time. It pulls current pages, reads them, and assembles an answer with citations. This is where a brand with no history can appear immediately, and where a well-known brand with a poorly structured site can lose to a smaller competitor whose page answers the question in a liftable form.
Two inputs, two different remedies. Priors respond to consistency and coverage over time. Retrieval responds to structure and freshness right now.
The spread
Because their retrieval logic and source preferences are genuinely different, and the spread is measurable.
SOCi's 2026 Local Visibility Index looked at more than 350,000 business locations across 2,751 brands and measured how often each engine actually names a local business. ChatGPT recommended 1.2 percent of locations. Perplexity recommended 7.4 percent. Gemini recommended 11 percent. Google's traditional three-pack surfaced a location 35.9 percent of the time.
| Engine | Leans on | What it rewards |
|---|---|---|
| ChatGPT | Model priors plus a blended retrieval stack, including its own crawler and index | Broad, consistent brand footprint and crawler access |
| Perplexity | Live retrieval, citations-first design | Clean, recently updated, clearly sourced pages |
| Gemini | Google infrastructure and corroborated sources | Structure, authority signals, third-party corroboration |
| Claude | Retrieval with weight on clarity and completeness | Well-structured, unambiguous, corroborated content |
| Google AI Overviews | Content already performing in Google search | Classic search fundamentals plus experience signals |
Same businesses, same category, four different answers. That is not noise. That is four systems applying four different thresholds for what counts as enough evidence to name a business.
A program built around one engine produces one engine's result. Test all four, because your buyers are not all using the same one.
Under the hood
No. BrightEdge found in February 2026 that only around 17 percent of AI Overview citations come from pages ranking in Google's organic top ten. Other 2026 studies put the overlap between 17 and 38 percent depending on methodology, down from roughly 76 percent measured by Ahrefs in mid-2025.
Authority metrics do not close it either. In the Semrush category analysis run by Kevin Indig across 1,094 US categories between January and June 2026, covering more than 50,000 brands and 600,000 citations, Authority Score predicted category ownership 52.5 percent of the time. Domain strength predicts an AI citation about as well as a coin. Indig's own conclusion: "traditional SEO metrics aren't enough to explain who owns a topic."
The commonly repeated version is that ChatGPT runs on Bing. That was accurate at launch and it has loosened since. Bing supplied the original index and still contributes to it. OpenAI subsequently began operating its own crawler and building its own index, and has signed direct content licensing deals with a long list of publishers whose material reaches ChatGPT through structured feeds rather than open-web crawling.
The defensible reading is a blended stack rather than a single upstream index. OpenAI has never published a full architecture diagram, and observed behavior has changed several times.
The published research disagrees on how much Bing rank matters, and the disagreement is instructive. One 2026 analysis found Bing's top-three URLs matched actual ChatGPT citations only 6.8 to 7.8 percent of the time. A separate Seer Interactive study of more than 500 citations put the match at 87 percent. Being indexed in Bing is clearly necessary. Ranking first in Bing is not the lever it gets sold as.
OpenAI operates three separately identified user agents with different jobs. They can be controlled independently in robots.txt, and treating them as one entity is the most common technical error in enterprise deployments.
| Agent | What it does | If you block it |
|---|---|---|
| GPTBot | Collects content that may be used to train future models | You withhold training data. Citation eligibility survives |
| OAI-SearchBot | Builds and refreshes the search index powering ChatGPT results and the citations attached to them | You lose citation eligibility. This is the one that matters |
| ChatGPT-User | Fetches a specific page in real time when a user or an in-conversation action requires it | Live reads stop mid-conversation |
This gives you a legitimate middle position. Blocking GPTBot while allowing OAI-SearchBot keeps your content out of model training and preserves your eligibility to be cited. If your legal team has opinions about training data, that is the configuration to hand them.
Blocking all three at once removes the site from the retrieval pipeline entirely and silently. Several bot management vendors did exactly that by default for a period, which means it may have happened to you without anyone deciding it.
One more practical detail. OpenAI's crawlers do not execute JavaScript. A March 2026 experiment confirmed ChatGPT parses HTML only. If your pricing, product names, or descriptions load after JavaScript runs, the crawler cannot see them, and what it cannot see it cannot cite.
The uncomfortable answer
Usually because your competitor has more third-party evidence, not a better website.
Muck Rack's December 2025 analysis, What Is AI Reading, found 82 percent of AI citations come from earned media. Omniscient Digital, reviewing more than 23,000 citations, found roughly 77 percent of the sources cited in answers about a brand are off-page.
So when a model is asked which vendor to use, it is mostly reading pages your competitor also did not write. Review sites. Roundups. Comparison articles. Podcast show notes. Trade coverage. Directory listings.
A competitor who has been covered, quoted, and listed in twenty places has twenty pieces of corroborating evidence. You have your own homepage insisting you are the best.
Models are built to weight corroboration. One source claiming something is weaker than five sources agreeing on it, and that is the correct behavior for an answer engine even when it is inconvenient for your brand.
Keep going
Which engines changed their answers this month, the prompt sets being run, and what actually moved a brand into the shortlist. Written by people running the prompts, not summarizing someone else's report.
No spam. Unsubscribe anytime.
What you control
Yes, through three levers, and none of them are prompt tricks.
Your brand name spelled identically everywhere: your site, LinkedIn, Google Business Profile, directory listings, review platforms, and press. Inconsistent naming splits your entity into fragments, and a fragmented entity accumulates weaker associations than a single consistent one.
A self-contained answer near the top of each page, 40 to 60 words, that makes complete sense with no surrounding context. Sections that stand alone. Headings phrased as questions. Unliftable content does not get cited regardless of quality.
Third parties writing about you in ways retrieval systems can find. Original data is the highest-yield version, because it gives other people a reason to reference you by name. Every widely quoted statistic in this category exists because an organization published research specifically to be cited.
What does not work: stuffing your site with your own superlatives, publishing more posts at the same shallow depth, or trying to instruct the model directly in your page copy.
Answer engines are corroboration machines. You persuade them the way you would persuade a careful researcher, by being findable, clear, and confirmed elsewhere.
Common questions
Sometimes. ChatGPT combines what it learned in training with live retrieval from a blended stack that includes Bing, its own index built by OAI-SearchBot, and licensed publisher feeds. Whether it reads your site at question time depends on the query and on your site being crawlable. Confirm OAI-SearchBot is not disallowed, since that is the agent that determines citation eligibility.
GPTBot collects content that may train future models. OAI-SearchBot builds the search index that powers ChatGPT results and the citations attached to them. ChatGPT-User fetches a specific page live during a conversation. They are controlled independently in robots.txt, and blocking GPTBot while allowing OAI-SearchBot withholds training data while preserving citation eligibility.
Because it is assembling a description from third-party sources, and roughly 77 percent of what it cites about a brand is off-page, per Omniscient Digital's review of more than 23,000 citations. Outdated directory listings, old press, and inconsistent naming produce inaccurate answers more often than a bad website does.
Test all four, then prioritize by where your buyers actually are. Engine behavior varies widely. SOCi's 2026 index found ChatGPT names 1.2 percent of local business locations while Gemini names 11 percent, so a single-engine program gives you a single engine's result.
Not in the organic citation itself. Advertising products inside AI surfaces are emerging separately, but a cited source in a generated answer is earned through retrieval and corroboration, not purchased. That is why citation supply and structure carry the weight here.
Faster than traditional search reporting assumes. AI content freshness cycles run on roughly a 70-day cadence, which is why monthly measurement is the correct rhythm for this work and quarterly reporting describes a state that has already moved.
Start here
Engine behavior varies enough that a single-engine check tells you almost nothing. The AI visibility audit from Conspicuouz Creative Group runs your category's real prompts across ChatGPT, Claude, Gemini, and Perplexity, records who gets named instead of you, and identifies the structural reason behind the pattern.