In short
- A source link identifies material associated with an answer, while a recommendation requires reading what the answer says about the business.
- Separate questions that name your brand from questions that ask the engine to find a suitable supplier.
- Check the documentation for each product because search access and source displays differ across engines.
Imagine a fictional bicycle shop whose maintenance guide appears beneath an AI answer. The answer recommends another shop for electric-bike repairs. Your guide received a citation. The other business received the recommendation.
Before rewriting the guide, establish what you wanted the answer to do. A citation can be useful evidence of source selection. It doesn't establish that the engine considers your business suitable for the customer's task.
Being retrieved, cited and recommended
We use the following distinctions when reviewing answers. They describe different observations, so keep separate columns for them.
| Observation | What we mean | What to record |
|---|---|---|
| Retrieved | Material was returned to the system through search or a page fetch. | The retrieval record, when available, and its URL. |
| Cited | The answer displays a source reference to a page. | The link and the statement it appears to support. |
| Mentioned | The answer names the brand. | The exact passage, including criticism or exclusions. |
| Recommended | The answer presents the brand as a suitable option for the user's need. | The recommendation and any conditions attached to it. |
This is our reporting convention. Avoid filling a missing retrieval field by inference. A server log records a request, but doesn't connect that request to a particular recommendation without additional evidence.
Source displays also need interpretation. OpenAI says ChatGPT's Sources panel may contain cited sources and other relevant links. Gemini likewise distinguishes sources from related content. 1 2
We therefore recommend saving the complete answer. Check the linked page against the statement beside the reference. A citation marker alone doesn't prove that the source supports every sentence in its paragraph.
For a worked diagnosis, read AI cites your website but doesn't recommend your brand.
Questions that already name you
Consider these fictional questions. "Does Willow Cycles repair electric-bike motors?" names the business. "Who repairs electric-bike motors near my office?" leaves the choice open.
We call the first a branded question. The second is unbranded with respect to Willow Cycles. Keep that distinction even when both questions concern the same service.
For the branded question, our recommended first check is factual accuracy. Does the shop clearly explain which repairs it offers? Are booking conditions current? Can a customer verify the answer against its public pages?
For the unbranded question, inspect the criteria in the request. Location matters in this example. The repair type matters too. A business that doesn't meet those constraints shouldn't be counted as a missed recommendation.
Use this separation when setting goals. Correcting an answer about your opening hours serves existing awareness. Being considered for a supplier shortlist serves discovery. Report those outcomes separately.
What the engines document
The descriptions below concern web search and source handling. They don't describe a universal formula for choosing brands. Apply each statement only to the product named.
ChatGPT Search
OpenAI says ChatGPT may search automatically when current information would help. It can rewrite a request into targeted queries for search providers. Follow-up searches may refine the results. 1
For publishers, OAI-SearchBot supports ChatGPT search inclusion. GPTBot gathers material that may be used for model training. Those controls are independent. 4
OpenAI says sites blocking OAI-SearchBot won't appear in search answers, although navigational links remain possible. ChatGPT-User handles certain user-initiated visits. It doesn't control search eligibility. 4
Our practical reading is to check the relevant search access before debating training permissions. Then inspect the actual answer. OpenAI warns that citations and results can be incomplete or incorrect. 1
Perplexity
Perplexity describes its answer engine as searching the open web and returning responses with citations. Its product overview also describes routing questions across multiple models. 5
Its crawler documentation assigns PerplexityBot to surfacing websites in search results. Perplexity-User can visit a page during a user's request. Neither is described as a foundation-model training crawler. 6
The same documentation addresses firewall access as well as robots.txt. It recommends matching bot identities against their published IP ranges. 6
Record the Perplexity mode used in an observation. Avoid treating the name of an underlying model as proof that another company's standalone app produced the answer.
Gemini
Google describes Gemini as combining its model with conversation context and external material, including Google Search. Uploaded files and connected services can also supply information. 7
Its help page says responses may show sources or related links. Some responses have no source links. Public websites and connected Workspace content are among the possible linked materials. 2
For a public visibility check, record whether connected material influenced the response. Keep Gemini observations separate from Google Search observations. The app's source display doesn't establish how AI Overviews selected its links.
Google AI Overviews
Google says AI Overviews may issue related searches across subtopics, a technique called query fan-out. An overview doesn't appear for every query. 3
Supporting pages must be indexed and eligible for a Search snippet. Google's separate Search Console control also governs inclusion in Search generative AI features. 3 8
That control can include or exclude a site's content and links. A child property may inherit its parent's choice. Google states that this setting doesn't control training. 8
Our recommendation is to verify eligibility before changing copy. Related searches also give a reason to inspect the whole buying question, including its constraints.
Microsoft Copilot
Microsoft describes consumer Copilot as searching the web for relevant content, then summarizing it with source links. It also warns that Copilot can misrepresent source material. 9
Microsoft's documentation for Copilot Chat identifies the Bing index as a source of public-web grounding. With work-data grounding enabled, Microsoft 365 information can also inform answers. 10
Keep the product and account context in your record. An answer drawing on internal company files isn't a clean observation of public brand discovery.
We recommend investigating both the cited page and the claim about your business. Source availability and the accuracy of the generated summary require separate checks.
Claude
Anthropic says Claude can invoke web search for questions needing current information. Search responses include citations. Web fetch can retrieve page content when the user supplies a URL. 11
Its publisher documentation separates Claude-SearchBot from Claude-User and ClaudeBot. Their stated roles are search indexing, user-directed retrieval and potential model training respectively. 12
Anthropic says blocking Claude-SearchBot can reduce search visibility. It also states that its bots honor robots.txt and don't bypass CAPTCHA protections. 12
In your observations, distinguish a page Claude found from a URL you supplied. Directly handing over the page demonstrates a different route from discovering it independently.
Turn a missing recommendation into a useful investigation
Return to the fictional bicycle shop. Suppose the answer cites its maintenance advice but excludes it from electric-bike repairs. Start with the service itself. Does the shop actually provide it?
If it does, compare the public service page with the answer. Identify the exact missing or incorrect statement. A product specialist should confirm the correction before it reaches the website.
If the answer links to another publisher's outdated description, prepare a correction supported by the shop's current information. Keep the external publisher's editorial decision separate from your own website changes.
When retrieval isn't observable, mark it unknown. When the answer names the shop only as an information source, record a mention. Reserve recommendation for a passage proposing it as a suitable choice.
Then decide which work the evidence supports. Access problems belong in technical foundations. Missing product facts belong in pages that answer for your brand.
Use measuring AI visibility to preserve the distinction in reporting. Keep the same question wording when comparing observations, and record changes to the product or search mode.
The next action should follow from the saved answer. "Our booking conditions are missing" is a fix someone can verify. "The AI doesn't trust us" leaves the cause untested.
Sources
- OpenAI, Searching the web with ChatGPTAccessed Back to the text, first mention of source 1
- Google Gemini Apps Help, View related sourcesAccessed Back to the text, first mention of source 2
- Google Search Central, AI features and your websiteAccessed Back to the text, first mention of source 3
- OpenAI, Overview of OpenAI CrawlersAccessed Back to the text, first mention of source 4
- Perplexity, Product overview and source explanationsAccessed Back to the text, first mention of source 5
- Perplexity, Perplexity CrawlersAccessed Back to the text, first mention of source 6
- Google, An overview of the Gemini appAccessed Back to the text, first mention of source 7
- Google Search Console Help, Search generative AI controlAccessed Back to the text, first mention of source 8
- Microsoft, How Copilot finds its sources of informationAccessed Back to the text, first mention of source 9
- Microsoft Support, Frequently asked questions about Microsoft Copilot ChatAccessed Back to the text, first mention of source 10
- Anthropic, Enable and use web searchAccessed Back to the text, first mention of source 11
- Anthropic, Web crawlers and site owner controlsAccessed Back to the text, first mention of source 12
