AI search platforms: why your visibility differs by product

What an AI search platform is, and what each product says about its sources

An AI search platform is a product that reads a question, decides whether to look things up, and writes an answer, usually with links to the pages it drew on. Google's AI Overviews and AI Mode, ChatGPT search, Perplexity, Microsoft Copilot, Claude and Brave's AI Answers all work this way. Marketers call the work of appearing in them answer engine optimisation (AEO), generative engine optimisation or AI SEO. Those are three names for one job, and what is AI SEO sets it out.

These platforms don't share one index or one set of rules. Each product decides when to search the web and which pages to show as sources, even when two products draw on the same index. Google's own documentation makes the point about two of its Search features: AI Overviews and AI Mode "may use different models and techniques, so the set of responses and links they show will vary." Products can also overlap underneath. OpenAI's help centre says ChatGPT search sometimes partners with other search providers, and it names Microsoft among them.

So AI visibility is really a stack of separate readings. First comes the product: ChatGPT, Google AI Mode, Perplexity, Copilot, Claude. Then comes the surface, meaning the way in: the app a customer opens, or the API (the direct connection software uses), which some tracking tools call. Then come the conditions: signed in or out, the plan, the location, the model. A score that blends them can move when nothing about your business has changed.

How to compare AI visibility reports covers the maths behind a blended score. This guide covers the products themselves, and how to keep them apart in your own testing.

The Gemini app has no row of its own in the table below. Gemini is really three things, the Gemini app, Google Search's AI features and the Gemini API, and Gemini business visibility explained takes them one at a time. Perplexity's crawler checks are in how to appear in Perplexity sources.

Brave is the odd one out. Brave says its search runs on its own index, and its crawler help page says that if Googlebot can't crawl a page, Brave's crawler won't crawl it either. Brave's crawler also doesn't announce a separate user agent, so you can't target it by name in robots.txt. Your Googlebot rules are the setting that counts there.

In the last column, neither Google's nor Microsoft's report gives ChatGPT, Perplexity or Claude a reading of their own. Google's covers only Google's generative AI features in Search and Discover. Microsoft's folds Copilot, Bing's AI summaries and unnamed partner integrations into one set of figures, so treat it as a single reading. Here's what the help pages and developer documentation said in September and October 2026. Products change often, so recheck the source before you act on a detail:

Local details, plans and paid placements change what a customer sees

For a local business, the most useful difference between products is where each one gets details such as your address, hours and service area. Each company documents this differently, and some don't document it at all, so the table sticks to what each one says. Where a cell says a source isn't named, test with the suburb written into the prompt and don't assume where the details came from.

The same product can also show different people different things depending on the account. A free account, a paid plan, a work account and a signed-out visit are four separate conditions, and each belongs on its own row in your log.

Paid placements are the third thing to watch. Google Ads Help says ads can appear above or below AI Overviews in every market where AI Overviews run, and in Australia also within them, in English. OpenAI's help centre says ads may appear for ChatGPT Free and Go users, below the end of a response, labelled as sponsored and visually separate, and that ads don't influence ChatGPT's answers. OpenAI started testing ads in the United States on 9 February 2026 and says it is rolling them out in select regions. Whichever product you test, a sponsored unit is a paid placement, so never count one as a citation, yours or a competitor's.

The app and the API are different products

The same brand can sit behind two very different products. The app is what your customers open. The API is how software sends the same company's models a question, and it's what some AI visibility tools use. In the API, the developer holds the settings that the app works out for itself.

Search often isn't automatic. OpenAI's docs say its Responses API model "can choose to search the web or not based on the content of the input prompt". Anthropic's docs say Claude determines when to search based on the prompt, and Google's Gemini API docs say the model decides whether a Google Search can improve the answer. A run with no search answers from what the model already knows, with no sources to cite.

Most other settings sit with the developer too. OpenAI lets a developer turn off live page fetching, so its tool answers from cached or indexed results only. Microsoft's Bing grounding tool returns five results by default, and Microsoft notes the model might not use all of them. None of these settings is wrong. They mean an API result describes that setup, and nothing wider. Here's how the main settings compare:

Live search or memory: read each answer as one draw

Before you read anything into an answer, check whether the product searched at all. An answer written from memory reflects what the model learned in training, so it can repeat an old phone number or miss a service you added last month, and no fix on your site will show up until a search runs. Then remember that a searched answer is one draw. The product picks its own searches each time, so the same question can return different sources an hour later.

Use these checks before you log a result:

• ChatGPT: OpenAI says responses that use web search may include citations, and that a Sources button, when available, lists the cited sources and other relevant links. If neither appears, log the row as "no search". To force a search, pick Search from the tools menu, or regenerate the answer and choose Search the web.
• Claude: Anthropic says you'll see an indicator while Claude searches the web, and that search answers include citations. An answer with no indicator and no citations came from memory.
• Google AI Mode: Google's help page says AI Mode may show a set of web links instead of an AI response when it doesn't have high enough confidence in the answer. Log that as its own outcome, separate from an answer that left you out.
• Repeat before you conclude. Google's AI Mode help suggests asking multiple versions of your question, so run each prompt several times on the same day and again on a later date, and judge the pattern across runs.
• Read a single miss as one draw. If you were left out once and named in the other runs, that's a normal spread between runs. If you're missing from every run in one product and present in another, compare the two products' access rules in the first table before you change any page.
• Change one thing at a time. If you edit a page, update robots.txt and claim a listing in the same week, a later change in the log can't be traced to any one of them.

Log each product and surface as its own row

Keep one log, with every test on its own row and every surface labelled. Then apply two rules. Compare a row only with rows that match on product, surface, account state, plan and location. And don't average app rows with API rows, because an API run tells you how one configuration answers while an app run tells you what one person saw. If a tracking tool can't say which surface it used, label its rows "surface unknown" and leave them out of comparisons.

The AI visibility checklist sets up a dated test log, with columns for the date, product, conditions, prompt and cited URLs. Keep those, and add the settings that tell one row from another, so anyone can rerun a test the same way:

• Surface and tool, written out in full, such as "ChatGPT app", "OpenAI API, web search tool" or "Microsoft Foundry, Grounding with Bing Search".
• Account state and plan: signed out, free, paid or a work account. Each can change the model on offer, the usage limits and, in ChatGPT, whether ads appear.
• Whether a search ran. A run with no search answered from what the model already knew, so keep it apart from runs that searched.
• The location the call sent, exactly as sent: fields such as city, region and country, or a Bing market code such as en-AU. Write "none" if the call sent nothing.
• Any domain filter, with the allowed or blocked list pasted in full.
• Whether live fetching was on. OpenAI's tool runs cache-only, answering from cached or indexed results, when external_web_access is set to false.
• Any result or search limit in the call. Bing's grounding tool returns 5 results unless told otherwise (up to 50), Perplexity's max_results takes 1 to 50, and Claude's max_uses caps the number of searches.
• How you appeared: named with a link to your site, named with no link, or not named. Note any sponsored units separately and never count them.
• For OpenAI's API, the full sources list beside the citations, since OpenAI says it's often longer and it shows pages the model consulted but didn't cite.

Questions to ask a tracking tool vendor

Before you trust a tracking tool's report, get a usable answer to each question below. Some tools also list products such as Meta AI, Grok or DeepSeek; ask the same questions about those rows, and leave out any row the vendor can't explain. Put these questions to the vendor:

A worked example: one prompt set, three products

Say an electrician in Caloundra asks the ChatGPT app on their phone, signed in, "who installs EV chargers in Caloundra?" Their business gets named. A week later a tracking report says the same question, sent through OpenAI's API, didn't name them. Search Console shows AI impressions for their EV charger page from Australia that month.

None of these contradicts the others. They're three products with three sets of settings. The business, which is hypothetical, settles it with a small fixed prompt set written the way customers ask, rerun the same way in each product. The steps below are illustrative, not a benchmark:

1. Write three prompts and keep them fixed: one service and suburb question ("who installs EV chargers in Caloundra?"), one problem-led question ("my switchboard trips when the EV charges, who can fix it near Caloundra?") and one comparison question ("what should I check before hiring an EV charger installer on the Sunshine Coast?").
2. Run each prompt in the ChatGPT app twice, once signed in and once signed out, and in Perplexity and Copilot signed out. Note on every row whether a search ran.
3. Check the API run's location. If none was sent, the search didn't know where the question came from, so it can't stand in for a Caloundra customer. Rerun it with the city set to Caloundra and the country set to AU, and log it as an API row.
4. Record how the business appeared each time: named with a link, named without one, or absent, plus any sponsored unit shown.
5. Keep the Google figures as their own trend line, read with how to interpret Google AI impressions.
6. Record the model name on every row, so a model change later doesn't look like a change in visibility.
7. Repeat the full set on a later date before deciding anything. If the business is named without a link in one product, check that its address and phone match everywhere that product reads them, using AI answers and accurate company facts, before touching the site.