How AI assembles
its answers

RAG stands for Retrieval Augmented Generation and is the technology that powers most AI search engines. The principle is straightforward: instead of relying solely on what the model learned during training, it retrieves current information from the web for every question and incorporates that into its answer.

Without RAG, ChatGPT could only answer based on its training data, which is always outdated. With RAG, it can pull in current information: recent reviews, current prices, new businesses that opened last month. This is the reason AI answers in 2026 are far more current and specific than they were two years ago.

For UK businesses, this has a direct implication. Your online presence is not just about Google rankings any more. Every page on your website, every Trustpilot review, every mention in a trade directory is potentially retrievable by AI. The question is whether AI can find it, understand it and use it.

How the retrieval process works

The retrieval process works like this: AI sends a search query to the web, receives a list of potentially relevant pages, reads those pages and extracts the most relevant passages. Those passages are offered as context to the language model, which processes them into its answer. The model effectively "reads" a selection of the internet before responding. The quality of this retrieval step directly determines the quality of the answer.

The selection of which pages to retrieve is not random. AI uses signals similar to search engine rankings: domain authority, page structure, freshness and relevance to the query. But there are differences. AI also considers whether a page contains quotable passages, whether the information is specific enough to cite and whether multiple sources confirm the same facts.

Research shows that 40% of sources cited in Google AI Overviews come from positions 11 to 20 in traditional search results. This means a page ranked 15th on Google can still appear in an AI-generated answer. Content quality and citeability matter more than your exact Google position. A well-structured page at position 12 gets cited whilst a vague page at position 3 gets skipped.

This has direct implications for how you structure your content. Your page is not read and understood as a whole. The system extracts passages. If your most important information is buried in a long paragraph halfway down the page, there is less chance it gets picked up than if it sits prominently at the top under a clear heading. Content structure determines whether AI finds the right information about your business.

Ask ChatGPT "best accountant in Manchester" twice, five minutes apart, and you will likely get different answers. This is not a bug. It is fundamental to how language models work. Fewer than 1% of ChatGPT citations remain consistent when the same question is asked repeatedly. The UK shows 60% citation volatility in AI results, higher than France at 42%.

Several factors cause this variation. First, the retrieval step may return different web pages each time, especially if the search index has been updated. Second, the language model has a "temperature" setting that introduces controlled randomness into word selection. This prevents AI from giving identical, robotic answers but means businesses cannot guarantee a permanent recommendation.

What this means for monitoring

Because answers vary between sessions, checking your AI visibility once tells you very little. You need to monitor consistently over time. A single query where your business appears does not mean you always appear. Equally, one session without a mention does not mean you are never recommended. VestVale tracks your AI visibility across multiple queries and platforms over time, giving you a reliable picture rather than a snapshot. This is why monitoring local AI visibility matters more than ever.

Cross-platform differences add another layer. ChatGPT, Gemini and Perplexity use different source indexes, different ranking algorithms and different language models. A business that appears consistently in ChatGPT might be absent from Gemini because Gemini pulls from Google's Knowledge Graph whilst ChatGPT uses its own web crawling. For a detailed comparison, see our guide on why AI results differ between platforms.

The practical takeaway is that AI visibility is not binary. Your business is not simply "visible" or "invisible." It exists on a spectrum, appearing in some percentage of relevant queries, on some platforms, for some variations of the question. A strong AI visibility strategy aims to increase that percentage across all platforms, not to guarantee a single result.

This is different from traditional SEO, where you can rank at position 3 for a keyword and stay there for weeks. In AI search, your position fluctuates with every query. The businesses that appear most consistently are those with the strongest, most diversified online presence: strong website content, consistent reviews, multiple directory listings and mentions across platforms. Consistency of information across these sources is what gives AI confidence to recommend you.

A fundamental difference between AI answers and Google search results is conversational context. With Google, every search is standalone. You type a keyword, you get results, done. With AI chat platforms like ChatGPT, there is conversation history. Earlier messages in the conversation influence the answer to your next question.

If someone first says "I run a limited company in Edinburgh with five employees" and then asks "Which accountant do you recommend?", AI factors in that context. The answer will include accountants who specialise in limited companies, preferably with experience handling payroll and HMRC compliance in Scotland. Without that context, AI would give a more generic list.

Specificity wins over generality

This has implications for how you build your online profile. The more specifically you describe who you serve and what you specialise in, the better AI can match you to specific user questions. An accountant whose website states "Specialist in limited companies with 1 to 10 employees, Edinburgh and the Lothians" will be recommended more often for exactly that audience than one who writes "We help businesses with their accounts."

The role of intent is growing as AI models become better at understanding what users actually want. According to Ofcom, 75% of UK adults have encountered AI-generated search summaries. As familiarity grows, so does trust, and users increasingly rely on AI to select the right sources and deliver the right answer. That trust makes AI's selection process even more consequential for businesses.

Question wording matters enormously. "Cheap solicitor London" produces a different answer from "Experienced solicitor London." AI interprets "cheap" as price-focused and highlights firms advertising fixed fees or free consultations. "Experienced" leads to firms with decades of practice and high review scores. Same profession, same city, but different firms in the answer.

Reviews as context for AI matching

For businesses, this means you should optimise not just for your core terms but also for the words your customers use to describe you. If your clients call you "thorough" and "professional" in Trustpilot reviews, those are the words AI picks up. If clients describe you as "quick" and "affordable," you get recommended for different questions. Your review language literally becomes the context AI uses to match you to user queries. This is why managing your online reputation is now directly tied to your AI visibility.