ilumi SEO

AI-generated misinformation can distort how products and brands are represented online. At the same time, guidance on AI Search remains fragmented and often contradictory. There is little official documentation from Google, OpenAI or other related platforms. Especially information explaining how businesses should optimise their websites to reduce AI hallucinations, correct inaccurate brand information, or increase the likelihood of being referenced by AI-powered search.

In response, a growing number of SEO platforms now offer AI visibility and citation monitoring tools. However, we found many of these solutions to be incomplete and difficult to validate. Today, there is no official method for measuring AI-driven brand visibility. In practice, what the SEO platforms are actually doing is tracking referral traffic and UTM parameters from links shared by tools such as ChatGPT and Claude. There are no other known solution of tracking 'AI visibility' hollistically and without UTMs.

The lack of clear guidance for website owners is further complicated by changes in how search engines discover and process content. Google's reduced crawl activity means many websites receive less attention than they did in the past. New content can take longer to be discovered, indexed, and associated with important keywords, making it more difficult for brands to influence both search and AI-generated responses.

Understanding how LLMs discover, process, and interpret website content allows organisations to optimise their digital presence for AI-powered search. This can improve the accuracy of AI-generated responses and reduce the likelihood of misinformation or brand misrepresentation.

How this works behind the scenes

LLMs do not process websites in the same way as traditional search engines. Rather than evaluating an entire page for ranking purposes, they retrieve specific content fragments that appear most relevant to a user's query. These fragments are then used to generate a response, summary, or citation.

During the webinar, we demonstrated this process using publicly available LLMs. We simulated real user queries from the UK and visualised exactly what information the models retrieved from a website before generating an answer.

When users ask a question, AI systems do not always rely on their training data alone. For many commercial and product-related queries, they actively retrieve information from websites in real time. Users may see messages such as "Searching for..." while the model gathers information from relevant sources. The model then selects the content it considers most useful and incorporates it into the final answer. If key information is missing or difficult to retrive for an LLM model, your website may be overlooked even when it contains the correct answer.

Throughout the webinar, we compared how AI systems interpreted website content before and after optimisation. The results showed that making content easier to parse helped AI consistently surface key product information and USPs for relevant user queries.

Deep Dive: How AI searches, retrieves and interprets information

In this section, we explored how AI systems retrieve and interpret information from websites. Through a series of simulations, we demonstrated that what users see on a webpage is not always what AI systems see when retrieving information.

AI models rely on a website's underlying structure to understand relationships between content, products, and key information. If these relationships are unclear, important details can be missed or interpreted incorrectly. Interactive elements, product filters, JavaScript-heavy content and complex page layouts can also make it harder for AI systems to retrieve accurate information.

We demonstrated how making content easier to parse helped AI systems better understand product details, USPs, specifications, and supporting information when generating responses.

The techniques we covered for closing this gap included:

Structured data that connects specifications, FAQs, and offers directly to the products or content they describe
Pre-rendered critical content
Clear internal linking between related pages and topics
Removing contradictory information, content duplication, and keyword cannibalisation
Using clear HTML structure to distinguish between similar products, services, and categories

How AI reponses can harm your brand

AI hallucinations may appear random, but they are often the result of lacking or 'difficult-to-retrieve' information. When AI systems cannot confidently identify the correct answer, they may fill gaps with incorrect associations.

During the session, we demonstrated how technical improvements can reduce these issues. Using a tyre manufacturer's website as an example, we analysed how AI systems interpreted product information, vehicle fitment data, and filtering options before and after a series of enhancements.

Tyre catalogues are particularly challenging because product specifications, fitment requirements, and filtering systems can be complex. We showed how relatively simple changes, including improved content accessibility and pre-rendering of key information, helped AI systems retrieve more accurate data and make better product recommendations.

The testing was conducted over a two-month period, tracking how AI responses evolved after the changes were implemented. Initially, AI systems frequently returned incomplete or incorrect recommendations. As the updated content became discoverable and easier to interpret, the correct tyre products were recommended more consistently. In some cases, the manufacturer also began appearing as a recommended brand for specific vehicle types because the AI could more easily connect the relevant product and fitment information.

To minimise bias, testing was carried out across multiple browsers and sessions using clean environments without personalised browsing data. This allowed us to focus on how the underlying website changes influenced AI retrieval and recommendations over time.

What the data says about search behaviour right now

Google updated its Search Central documentation in early February 2026 to confirm Googlebot now crawls only the first 2MB of a page's resources, down from 15MB - an 86.7% reduction in what gets crawled per resource. Pages and scripts that exceed that limit get truncated or skipped entirely.
AI Overviews now appear on roughly 48% of Google search queries as of March 2026, up from around 34.5% in December 2025 - a near-1.5x increase in three months.

Put together: less of your website gets crawled, more queries get intercepted by an AI summary before a click-through happens. Also being cited inside that summary is becoming the only way to capture attention at all.

Our own findings: most "AI-ready" sites aren't

We audited enterprise websites across the UK spanning automotive, retail, and fashion. The pattern was consistent: roughly 80% of the sites we reviewed stored business-critical content (FAQs, product specifications, fitment and compatibility data) inside components that AI systems couldn't parse.

The downstream effect was always the same shape: either the AI system skipped the content and recommended a competitor with cleaner structure, or it filled the gap itself and generated an inaccurate answer attributed to the brand. In every case, the content existed. It just wasn't usable by the systems now standing between the brand and the customer.

What this means for your digital marketing strategy

During the webinar, we shared primary research on how this works in depth and the impact it has across more than 100 websites and real user queries. We compared Gemini, Claude, and ChatGPT, analysing how each retrieves and uses information, and showed how the websites we worked on performed before and after optimisation.

For SEO professionals and marketing leaders, this is a challenging period where misinformation is often louder than reliable guidance. It is increasingly important to understand the technical foundations of how AI systems retrieve and interpret information in order to make better commercial decisions.

Based on the data we shared, the strongest results were achieved through improvements in content structure, accessibility and how clearly information could be interpreted by AI systems.

Free consultation