Why AI chatbots don’t always search the web – and why that’s a reputation issue

Most people now understand that AI chatbots can search the web.

What is less well understood is that they do not always search, they do not all search in the same way, and when they do search, they are not simply reading “the web” in some neutral or comprehensive sense.

This means that when customers, journalists, procurement teams or marketing leaders use AI tools to ask who they should consider, which firms lead a category, or which agencies are known for a particular type of work, the answer may depend less on an objective reading of the market and more on what the model remembers, retrieves, ranks and summarises.

In that context, AI visibility becomes part of the reputation management challenge, not just a technical search issue.

Let's unpack it a bit.

What do we mean by "grounding"?

Large language models are often described as if they “know” things. In practice, they generate answers based on patterns learned during training, plus any additional context they are given at the time of the query.

Grounding is the process of connecting the model’s answer to external evidence.

That evidence might be a document, a database, a company knowledge base, a search index, a set of uploaded files, or live web results. The purpose is to make the answer less dependent on model memory alone and more connected to current, verifiable information.

The underlying idea is closely tied to retrieval-augmented generation, usually shortened to RAG. This 2020 RAG paper describes systems that combine a model’s internal, learned knowledge with retrieved external information, rather than relying only on what is stored in the model’s parameters.

In simple terms: instead of asking the model to answer from memory, you give it material to work with.

That has become increasingly important as AI chatbots have moved from novelty tools to decision-support systems. A model that was trained months ago may not know about a new product, a recent award win, a merger, a reputational issue, a leadership change or a newly published ranking. Grounding is one way to bridge that gap.

What is web-search grounding?

Web-search grounding is a specific form of grounding where the chatbot uses web search results to support its answer. But it is not the same as typing a query into Google and reading the first ten blue links.

In practice, the process usually looks more like this:

1. A user asks a question.
2. The AI system interprets what the user is asking.
3. It may rewrite that question into one or more search queries.
4. It sends those queries to a search or retrieval system.
5. It receives candidate sources.
6. It selects, extracts and summarises information from those sources.
7. It blends that retrieved material with the model’s existing knowledge to produce a conversational answer.

OpenAI says ChatGPT Search may rewrite a user’s question into one or more targeted queries sent to search providers. OpenAI also says ChatGPT can choose to search based on what the user asks, or the user can manually select web search.

Google describes grounding with Google Search as connecting Gemini responses to search results based on publicly available web data.

Perplexity describes its system as searching the internet in real time, gathering information from sources and distilling it into a conversational answer.

So, when we talk about web-grounded AI answers, we are not talking about a chatbot simply “checking the internet”. We are talking about a retrieval and synthesis process that connects the user’s question and the final answer.

And that process is often where visibility is won or lost.

Why web search is not automatic, universal or neutral

Many people assume that modern AI chatbots automatically search the web whenever a question might benefit from current information, but that's not the case.

Web search is not automatic

Some AI answers are generated from model memory. Others are generated with live retrieval. The user may not always know which mode is being used, and the behaviour can depend on the product, settings, prompt, model and system design. A company, brand or agency can be “known” to the model because it has been heavily represented in training data. Or it can be retrieved because it appears in relevant, accessible, high-ranking, current sources. Those are different forms of visibility.

Web search is not universal

Different AI platforms search differently. ChatGPT, Gemini and Perplexity do not use identical retrieval systems, ranking signals, source selection rules or answer formats. A brand that appears prominently in one platform may be weaker in another.

Web search is not neutral

When an AI system searches, it has to make choices. It has to decide what the user really means. It has to choose search terms. It has to decide which sources are relevant. It has to extract some information and ignore the rest. It has to decide how much weight to give to model memory versus retrieved evidence.

For example, a user might ask: “Which PR agencies should we consider for a fintech launch?”

The AI might search for:

• “best fintech PR agencies UK” OR
• “financial services PR agencies London” OR
• “PR Week financial services agency awards” OR
• “UK technology PR agencies fintech clients”

Each of those searches could surface a different map of the market. The answer is shaped not just by the agency’s reputation, but by the AI system’s interpretation of the brief.

How grounding changes visibility in AI answers

Some brands are remembered. They appear when the model answers from training data alone. These are the names that have become embedded in the model’s representation of a category.

Some brands are retrieved. They appear when live search is enabled. These are the names that are discoverable through current web sources: rankings, awards, media coverage, directories, thought leadership, client pages, sector pages, Wikipedia, analyst mentions and other accessible sources.

Some brands are weak in both. They may be well known to people in the market, but not strongly represented in either the model’s memory or the sources that the model retrieves.

A model-memory answer may favour long-established names, widely discussed brands and legacy category leaders. A web-grounded answer may favour sources that are recent, well structured, highly ranked, easy to summarise or repeated across multiple pages.

Neither is a perfect representation of the market. They are just different visibility environments.

Example: UK PR agencies

I recently tested this using PRovoke Media's ranking of Top 70 UK PR Agencies.

The analysis covered 900 AI-generated answers across ChatGPT, Gemini and Perplexity, using prompts designed to simulate buyers looking for PR agency recommendations across different sectors.

The overall ranking was not the most interesting thing. The more useful finding was the difference between agencies that were stronger in model-memory answers and those that became more visible when live web search was enabled.

For example:

• Hope&Glory PR appeared far more often in non-web answers than in web-grounded answers: 132 mentions without web search versus 20 with web search.
• Weber Shandwick showed a similar pattern: 107 non-web mentions versus 41 web-search mentions.
• Edelman was strong across both, but still appeared more often in non-web answers: 236 without web search versus 136 with web search.

By contrast, some agencies became more visible when search was enabled.

• MHP Group had 48 non-web mentions versus 90 web-search mentions.
• The PHA Group had 24 non-web mentions versus 54 web-search mentions.
• FINN Partners UK had 3 non-web mentions versus 22 web-search mentions.

This data shows different types of AI visibility. Some agencies are embedded in model memory. Some are being retrieved through live search. Some are visible in both. Some barely appear at all. The same market looks different depending on whether the model is answering from memory or retrieving live evidence.

What this means for AEO and GEO

Answer Engine Optimisation (AEO) and Generative Engine Optimisation (GEO) are often described as just new versions of SEO, and yes, traditional search visibility still matters. When AI systems retrieve from web sources, then rankings, crawlability, page structure, schema, authority and content relevance all have an impact.

But AI visibility is also about shaping the evidence environment that AI systems use to answer category questions.

For agencies advising clients, that means asking different questions:

• Are clients visible for the categories they want to lead?
• Are they associated with the right capabilities, values and proof points?
• Do third-party sources describe them in the way they want the market to understand them?
• Does their owned content give AI systems clear, current evidence to retrieve?
• Are their awards, case studies, executive views and specialist credentials easy for AI systems to summarise?
• Do they appear in broad category questions, or only when searched by name?

And ultimately: “How do we make sure AI systems have the right evidence to include them in the right answers?”

What grounding means for PR and reputation

Grounding is becoming part of the AI reputation management challenge.

If an AI answer comes from model memory, the challenge is whether the client is already associated with the right category, strengths and proof points.

If an AI answer comes from web search, the challenge is whether the retrievable evidence supports the client’s current positioning.

One is about entrenched reputation and category association. The other is about the quality, accessibility and authority of the public evidence layer.

Agencies that understand this difference can give clients a more useful answer than “you need better AI visibility”. They can help clients understand not only whether they appear in AI answers, but why they appear, what evidence is shaping that visibility, and how that evidence can be strengthened.