Somebody has asked you to report on "AI visibility" and you now have four names on a whiteboard: ChatGPT, Perplexity, Gemini, Copilot. The practical question underneath is budgetary. Are these one channel you can sample once, or four surfaces that each need their own prompt runs, their own baseline, and their own row in the report?
The honest answer is four, but not for the reasons most comparison posts give. The differences that matter are not personality or answer quality. They are differences in where the text in the answer comes from, and those differences change what your measurement is actually measuring.
How To Read This Comparison (And Its Expiry Date)
This is a comparison of documented and observable sourcing behavior, not a quality ranking. None of these assistants is better than another at recommending brands, and any post that tells you one is has confused its own sample with a property of the system.
Three things you should hold in mind before you read the table.
Every row is a snapshot. Retrieval pipelines, citation displays, and default model versions change without much notice, and providers do not always document the change on the day it ships. A behavior that is true of a product in one month may be false the next.
"Assistant" is not "model." Each of these names covers a bundle: one or more underlying models, a retrieval layer, a ranking or re-ranking step over retrieved documents, a set of product-level rules, and a user interface. When your answer changes, any layer in that stack could be responsible. This is also why the same underlying model can behave differently in two products.
Free tier, paid tier, and API are not the same surface. The consumer product a colleague screenshots on their phone and the API endpoint your measurement runs against can differ in model version, retrieval access, and system instructions. Measure through official APIs and be explicit in your report about which surface you sampled.
Provider behavior changes without announcement. If you are reading this more than three months after the check date at the end of the post, treat every behavioral claim as unconfirmed and go to the primary docs.
The Comparison Table
Columns below are the stable comparison columns this site uses for assistant surfaces. Every row is a general characterization, not a measurement: it describes the shape of each surface as of the check date at the end of this post, so confirm a row against the current provider documentation before you rely on it. Do not restate any of it as settled fact on the strength of this table.
| Surface | How it assembles an answer | How citations are shown | Freshness posture | What a brand can plausibly influence | What this means for measurement |
|---|---|---|---|---|---|
| ChatGPT | Blends model knowledge with retrieval, and does not always retrieve. A conversational answer may be produced with no live lookup at all. | Links appear when retrieval fired; absent otherwise. Presence of a link is a signal that a fetch happened, not proof of what drove the wording. | Mixed. Some answers reflect training data of unknown vintage, others reflect a live fetch. | Being described consistently across the wider web; being present in the pages retrieval tends to reach. | You must record whether retrieval fired for each run. Mention rate with and without retrieval are different numbers and should not be averaged together. |
| Perplexity | Retrieval-first by design. The product is built around fetching sources and writing over them. | Citations shown inline and prominently as part of the core interface. | Skewed recent, because live fetching is the default path. | Being in the set of pages the retrieval layer surfaces for buyer-language queries. | Citation share is meaningful here because citations are near-universal. The same metric is much noisier on a surface that only sometimes cites. |
| Gemini | Google-family retrieval plus model knowledge, with grounding available and applied variably by surface and mode. | Grounding links and source panels appear in some modes and not others. | Mixed, and mode-dependent. | Presence and clarity in the sources its grounding reaches; consistent entity description. | Record the mode and the surface, not just "Gemini." Two runs labelled Gemini can be two different pipelines. |
| Copilot | Microsoft's assistant layer over search-backed retrieval, with behavior differing across consumer and work contexts. | Numbered source references in the response, availability varying by surface. | Search-backed and therefore skewed recent, subject to index coverage. | Coverage in the sources its retrieval draws on; entity consistency across those sources. | Enterprise and consumer contexts should never be pooled into one figure. Say which one you sampled. |
Two cautions about the table itself. First, the "what a brand can plausibly influence" column is deliberately vague, because the honest state of knowledge is vague. Nobody outside these companies has a validated causal model linking a specific publishing action to a specific change in answer composition. Second, none of these cells should be read as a promise. You cannot control any of these systems' outputs, and no method described on this site changes that.
Where The Four Genuinely Differ
Strip out the marketing and three real differences survive.
Whether retrieval happens at all
This is the largest difference and the one most measurement programs get wrong. A retrieval-first product will nearly always go and fetch something. A general assistant may answer from model knowledge with no fetch, especially for a broad question phrased in a way that does not look like it needs current information. The consequence is direct. If your prompt set is mostly broad category questions ("what's a good tool for X"), you will trigger retrieval at very different rates across the four surfaces, and your citation-based metrics will diverge for reasons that have nothing to do with your brand's standing.
What a citation means
On a surface where citations are the product, a citation is close to a complete account of where the answer came from. On a surface where retrieval is occasional, a citation tells you a fetch happened and a link was surfaced. It does not tell you how much of the answer's substance came from that page versus from model knowledge. Treating those two kinds of citation as the same data type is the single most common way a cross-assistant dashboard becomes misleading. See the companion piece on which metric answers which question for how the denominator changes underneath you.
How much the surface fragments
ChatGPT and Perplexity are relatively easy to name. Gemini and Copilot both span multiple modes and product contexts that can differ in retrieval and grounding. When your report says "Gemini," a reader is entitled to ask which Gemini, and you need an answer in your run log.
And a note on a fifth thing that keeps getting pulled into these comparisons: Google's AI Overviews are a search-results feature, not a conversational assistant, and they behave differently enough that mixing them in corrupts the comparison. That is a separate surface with separate measurement rules, covered in AI Overviews are a different surface.
Where The Differences Are Smaller Than Vendors Suggest
Some of the differences you will hear about are real but second-order.
Underlying model families overlap and rotate. Products change default models. Two assistants can sit on similar capabilities at the same moment. The retrieval and product layers usually explain more of the difference in output than the model does.
All four vary run to run. Non-determinism is not a property of one product. Every one of these will give you different answers to the same prompt across runs, which is why a screenshot from any of them proves nothing on its own. That mechanism is unpacked in why answers vary between people.
None of them is a ranking engine. There is no position to occupy. Being named first in a list of four is weak evidence about anything, on all four surfaces equally.
All four respond to the same underlying reality. Whether your brand is described clearly and consistently in the public record matters everywhere. The mechanisms differ; the direction of the requirement does not. The pillar overview of how ChatGPT decides which brands to recommend covers the shared substrate in more depth.
What This Means For Your Prompt Set And Your Reporting
Three concrete consequences.
Run the same prompt set against every surface you report on. If you vary the prompts by surface, you have lost the ability to compare. Same prompts, same period, separate columns.
Log the surface, the mode, the model version where the API exposes it, the date, and whether sources were returned. Those five fields are what turn a number into evidence. Without them you have an anecdote in a table.
Report per-surface. Never publish a blended "AI visibility score" across all four. A blended number hides the thing you most need to see: that you might be well represented on a retrieval-first surface and absent on a knowledge-heavy one, which points at completely different work. If a dashboard hands you one composite figure, ask what the weights are and how they were chosen. There is no principled universal weighting, only a choice someone made.
A reasonable starting shape for a first report: one row per surface, one column per metric, one clearly stated sample size per cell, and a footnote naming the run dates. Boring, defensible, and it will survive the first sceptical question in the room.
FAQ
Do I need to track all four? Only if your buyers plausibly use all four. Pick the surfaces your audience actually touches, say why you picked them in the report, and revisit that choice rather than quietly expanding forever.
Can I just take screenshots from each one? No. A screenshot is a single sample from a variable system, captured on an unrecorded date under unrecorded conditions. Use official APIs, record your run metadata, and keep screenshots for illustration only.
Which one should I optimise for first? That framing does not survive contact with the mechanics. There is no per-assistant optimisation dial. The work that plausibly helps is largely shared across surfaces, and where it differs it differs in emphasis rather than in kind.
Why is so much of this hedged? Because provider behavior changes and we will not state as fact what we have not sourced to a dated document or measured ourselves. That is the standard on every page here, and it is the reason this comparison describes patterns rather than confident specifics.
In Plain English
These four assistants are not one channel. The biggest real difference is whether the product goes and fetches sources every time or only sometimes, and that difference alone changes what a citation-based metric means on each surface.
This matters most to whoever has to build the first AI-visibility report and defend it, and to anyone about to average four surfaces into one number. Use this comparison when you are deciding which surfaces to sample and how to structure the columns. Do not use it as a ranking of the assistants, and do not rely on any row of it without checking it against current provider documentation.
Next: choose your surfaces, run one prompt set across all of them, log the metadata, and keep the results in separate columns. The AI search fundamentals hub covers the mechanisms behind each surface if you want the deeper version first.
Behavior on all four surfaces can shift without announcement. Use the briefing template for a dated summary of what changed and what it means for an existing measurement.
Primary-source verification
Checked July 28, 2026. The operational and factual guidance above was reviewed against the primary references below. Provider features, prices, policies, interfaces, jurisdictions, and account-specific results can change; recheck the controlling first-party document and your own account or measured data before acting.
- Google Search: AI features and your website - How Google describes eligibility, preview controls, and appearance in its AI search features.
- Google: common crawlers and special-case user agents - Current Googlebot and Google-Extended identities and controls.
- OpenAI: overview of OpenAI crawlers - Current GPTBot, OAI-SearchBot, and ChatGPT-User descriptions and publisher controls.
- Anthropic: web-crawler controls - Current ClaudeBot, Claude-SearchBot, and Claude-User descriptions and controls.
- Perplexity: crawler and user-agent documentation - Current PerplexityBot and Perplexity-User descriptions and controls.