Someone screenshots a summary at the top of a Google results page, someone else screenshots a chat reply, and both get pasted into the same Slack thread under the heading "we're not showing up in AI." Then a dashboard averages them into a single percentage and an executive asks what it means. The honest answer is that the number describes two different systems answering two different kinds of question, and the average belongs to neither.
This post separates them. Not as a taxonomy exercise, but because the separation changes what you measure, how often, and what you are allowed to conclude when one moves and the other does not.
A note on dating before anything else: the behavior of search-embedded summaries changes frequently, and specific claims about them go stale fast. Where this post describes how a surface behaves, it was checked against the provider documentation listed at the end, on the date given there; confirm it against the current documentation or a dated run of your own before you rely on it. If you repeat any of it in a report, attach the date you checked it.
The two surfaces, defined
A search-embedded summary is generated inside a search results page, in response to a query the user typed into a search box, and sits alongside the conventional results. Google's AI Overviews are the most visible example, and similar summary units appear in other search engines. The user did not ask for a conversation. They ran a search and received a synthesized paragraph or two plus links, above or beside the blue links they came for.
A conversational assistant is a chat product: ChatGPT, Claude, Gemini's assistant interface, Copilot, Perplexity. The user opens a session with the explicit intent of asking something, often in full sentences, often with follow-ups. The session carries context. Depending on the product and the settings, it may consult the live web, may rely on model weights alone, or may do both within a single answer.
The confusion is understandable, because these products increasingly share underlying models and both produce prose with citations. But the container matters more than the model. The container determines who arrives, what they typed, what the system decided to do about it, and what the user does next.
| Search-embedded summary | Conversational assistant | |
|---|---|---|
| Entry point | Search box, one query | Chat session, often multi-turn |
| Typical input | Short keyword-shaped query | Longer natural-language request, sometimes with constraints |
| Triggering | System decides whether to generate a summary for that query at all | An answer is generated for essentially every prompt |
| Context carried | Little to none between queries | Session history, and in some products stored memory |
| Competing content on screen | Conventional results, ads, other search features | Usually the answer alone |
| Next action | Often a click into a result, or no click | Often a follow-up question |
| What varies the output | Query wording, region, personalization signals, whether the unit fires at all | Prompt wording, session context, model version, retrieval settings |
What triggers each one
An assistant answers because you asked. The trigger is trivial: you sent a prompt, you get a response. The interesting variance lives downstream, in which sources get retrieved and which names survive into the final text.
A search-embedded summary has a prior step that assistants do not have. The system first decides whether to show a summary for this query, for this user, in this place, at this moment. That gate is a measurement problem all by itself, because a query that produces no summary is not a query you lost. It is a query where the surface did not exist. Averaging those into a visibility rate quietly understates you, or overstates a competitor, depending on how your tool handles the null case.
Availability and behavior also vary by region and by query type. The same query can produce a summary in one country and not another, and query categories that touch health, finance, or other sensitive areas are commonly treated differently by search systems. If you operate in several markets, a single global summary-visibility number is an average across systems that are not behaving identically.
Three consequences follow immediately:
- Record trigger rate as its own metric. What share of your prompt set produced a summary at all, in each market you care about?
- Report visibility conditional on trigger. "We appeared in 40% of the queries where a summary was generated" is a defensible sentence. "We appeared in 40% of AI results" is not, unless you define the denominator.
- Treat a fall in trigger rate as a separate event from a fall in inclusion. They have different causes and different responses.
Why the same brand can win one and lose the other
Once you accept that the two surfaces take different inputs, the divergence stops being a paradox.
Different query shape pulls different sources. A short search query and a paragraph-long chat prompt describing a team size, a budget constraint, and an existing tool are not the same request. The second one narrows the field before any retrieval happens. A brand that fits a narrow constraint well can be a reliable name in assistant answers while rarely surfacing for the broad keyword-shaped query that triggers a summary.
Different source pools. Search-embedded summaries are generated in a system with a full search index behind them. Assistants vary widely: some retrieve live, some answer from training data, some blend the two and the blend depends on the question. A brand that is well represented in recent web content but sparse in the material a model was trained on can win on one side and lose on the other. Our comparison of how each assistant sources an answer goes through those differences per product.
Different competitive field. In a search results page, the summary competes for attention with everything else on the page. Being absent from the summary while ranking well conventionally is a different commercial position than being absent from an assistant answer that fills the whole screen.
Different user intent. People arriving through a search box are frequently mid-research and will click. People in a chat session more often want the answer itself and stop there. Winning a mention in a surface where nobody clicks is worth something different from winning one where they do, and your reporting should not pretend the two are interchangeable units.
What breaks when you blend them into one score
Blending is tempting because executives want one number. Here is what the blended number costs you.
You cannot diagnose a move. A blended score drops four points. Did the summary trigger rate fall in one market? Did an assistant's model version change? Did a competitor publish something that entered a retrieval pool? A single figure gives you no way to ask the next question, so the investigation starts from zero every month.
The denominator becomes undefined. Assistants answer every prompt; summaries do not always appear. Combining a metric with a conditional denominator and one with a universal denominator gives an arithmetic result and no meaning.
Weighting happens by accident. If your prompt set has 60 assistant prompts and 20 search queries, your "AI visibility" number is mostly an assistant number wearing a broader label. Nobody chose that weighting. The set shape chose it.
Regional variation gets averaged into invisibility. A strong result in one market and an absent surface in another produce a mid-range blended figure that describes no market you actually sell in.
The wrong playbook gets funded. The responses differ. Assistant visibility work concentrates on being present, consistent, and clearly described in the sources those systems draw on, and on entity clarity so a model can tell who you are. Search-summary work overlaps with conventional search practice: being retrievable, being a credible result for the query in the first place, and having pages a system can quote cleanly. Give a team one number and they will pick one playbook, usually the one they already know.
How to report them side by side
Keep them in separate columns, permanently, with a shared prompt-set spine where the questions genuinely overlap.
A workable reporting shape:
- Segment the set. Tag every prompt as search-query-shaped or assistant-shaped. Some intents deserve one of each. Report both, and do not average them.
- Report trigger rate first for the search surface, split by market. Everything else on that side is conditional on it.
- Report inclusion and framing separately on both sides. Being named is not the same as being recommended, which is why the same prompt can look like a win in a count and a loss in the text.
- Attach provenance to every cell. Model or surface, version where knowable, run dates, region, sample size, and the prompt-set version. A cell without provenance is a rumor.
- Show variance, not just the point estimate. Both surfaces are non-deterministic in practice; a single-run figure will move on re-run for reasons that have nothing to do with your market.
- Date the behavioral footnotes. If your report explains why a surface behaves a certain way, that explanation carries the date it was verified.
Which figures belong in which column is a metric-selection question, and which metric answers which question walks through the options. The mechanics of running the measurement live in our guide to how to measure AI visibility, and the fundamentals of how each surface assembles an answer sit on the AI search fundamentals hub.
Do not let a blended number survive into an executive deck because splitting it is inconvenient. If someone insists on one figure, give them the two figures and a sentence saying which one is closer to the decision they are about to make.
Neither surface is on a path to replace the other, and this post makes no prediction about which grows. Both exist now, buyers use both, and a measurement programme that covers one and labels it "AI" is reporting on half its own scope.
FAQ
Is a search-embedded summary just an assistant with a different interface? It shares model technology in some cases, but it takes a different input, has a trigger gate, and sits in a page full of competing results. Those differences are enough to change what a visibility number means.
Should I run the same prompts on both? Run the same intents on both, in the phrasing native to each surface. Search queries are short; assistant prompts carry constraints. Forcing identical strings across both measures phrasing artifacts more than anything else.
How often should each be re-run? More often for the search surface, because it changes and because trigger rate is volatile. Set a fixed cadence for both and keep it, so that changes in your data reflect the world rather than your schedule.
What if my tool only reports one blended number? Ask the vendor for the segmentation and the denominator definition. If a tool cannot tell you what share of queries produced a summary at all, you cannot interpret its percentage.
In Plain English
Search-embedded summaries and conversational assistants are two surfaces, not one. The summary has a gate: it may not appear at all, and whether it appears varies by region and query type. The assistant always answers. That single difference makes the denominators incompatible, which makes any blended "AI visibility" percentage uninterpretable.
This matters most to whoever has to explain a monthly number: in-house search leads, agency strategists, and analytics teams building the report. Split the surfaces the first time you build the set, because retrofitting the split means restating past numbers.
Next step: tag your prompt set by surface, add trigger rate as a metric for the search side, and put the two side by side in your next report with dates and sample sizes attached.
Run the free AI Visibility Check to see the two surfaces reported separately, with the confidence range stated up front.
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.
- Apache HTTP Server: mod_log_config - Authoritative Apache access-log fields and formats.
- nginx: HTTP log module - Authoritative nginx access-log configuration and variables.
- Google Analytics: traffic-source dimensions - Current GA4 source, medium, campaign, and default-channel definitions.