Brand mentions in AI search and LLM answers: measure and improve

0:00 Brand mentions in AI search

A potential customer opens an AI assistant and types: which payroll software is best for a small business in Saudi Arabia? The answer names three companies. Yours isn't one of them. The customer never visits a search results page, never sees your ad, never reads your blog. You just weren't in the answer. This is happening in every category right now. In this lecture, you'll learn how AI search and assistants assemble answers, what Google has actually said about AI features, what really moves brand visibility in AI answers, and how to measure it properly with a sampling panel and a short Python script.

0:45 Why it matters for PR

Why does this matter for digital PR? Because AI answers draw heavily on what reputable, independent sources say about brands. The same coverage, reviews, expert commentary and industry comparisons that digital PR has always chased are now shaping which brands get named and how they're described. Think of AI assistants like a well-read friend giving recommendations. They've read the newspapers, the trade magazines, the review sites and the forums. They'll recommend whoever those sources talk about most clearly and most credibly. Your job is to be well represented in what they've read.

1:25 How answers are assembled

How are answers assembled? Simplified. Google's AI Overviews and AI Mode are built on Google Search's ranking and quality systems, and show supporting links to indexed pages. ChatGPT with search, Perplexity, Microsoft Copilot and Gemini combine the model's training with live web retrieval for many questions, often citing sources. And assistants without browsing rely on training data, which reflects what was widely written about you before the model's cutoff. Two consequences. First, independent sources matter enormously. Second, answers vary between runs, users, locations and products. So a single screenshot proves very little.

2:05 What Google has said

What has Google actually said? Its documentation on AI features says there are no additional requirements or special optimizations to appear in AI Overviews or AI Mode, beyond being indexed and eligible to show with a snippet. Normal SEO best practices apply. In May twenty twenty-six, Google published a guide to optimizing for generative AI features that repeats this. It says generative features are rooted in core ranking and quality systems, that llms dot txt files get no special treatment, and that you don't need to chop content into artificial chunks. Then in June twenty twenty-six, Search Console began showing generative AI performance reports, rolled out to all websites by the end of August.

2:55 Measurement options

Those Search Console reports show impressions for AI Overviews, AI Mode and generative features in Discover, broken down by page, country, device and date. At launch, they didn't include clicks or query data, so they tell you where you appear, but not for which questions or whether people clicked. They're evolving, so check the current documentation. For other AI products, you need your own measurement. Commercial tools can help. Ahrefs Brand Radar and Semrush's AI visibility toolkit track mentions and citations across several AI products. But check that they cover your markets and languages before relying on them.

3:38 What moves AI visibility

What actually moves AI visibility? Five things, and none of them are tricks. One, be clearly identifiable: canonical name, about page, Organization markup, consistent profiles. Two, be described by independent, reputable sources: coverage, trade press, expert commentary, reviews, and comparisons written by others. Three, publish specific, citable content: original data, clear pricing and product facts, methodology. Four, fix real reputation problems, because AI summaries often reflect recurring complaints. And five, keep facts consistent everywhere. And avoid manipulation, like mass-publishing self-promotional best-of lists on low-quality sites, fake reviews, or hidden instructions to AI systems. They risk spam policy violations and tend not to last.

4:23 Worked example 1: Lahore hotel

Worked example one, simple. A family-run hotel in Lahore asks an AI assistant, is this hotel good for families? The answer mentions an old complaint about parking and an outdated phone number. The owner checks where that came from: an old directory listing and several reviews from two years ago. She updates the directory and her Google Business Profile, publishes a clear parking and family facilities page, responds publicly to the old reviews explaining what's changed, and starts a compliant review program. Over the next few months, sampled answers mention the new family rooms and current contact details.

5:06 Worked example 2: payroll startup (illustrative)

Worked example two, a business scenario with illustrative details. A payroll software startup in Riyadh samples thirty prompts monthly across four AI products. Month one, it appears in few category answers, while two competitors appear often, usually citing a regional tech publication and a software review site. Brand prompts return outdated pricing from an old directory. So the team corrects the directory, publishes clear pricing and a salary data report with a real methodology, earns coverage in that regional tech publication, and runs a compliant review program. Over two quarters, its mention rate in category prompts rises, and brand answers show current pricing.

5:51 Watch me do it: the sampling panel

Watch me do it. First, prompts. I write thirty realistic questions customers would ask, in three groups: category, like best payroll software for small businesses in the UAE; comparison, like brand X versus brand Y for freelancers; and brand, like is brand X reliable? Some in Arabic too. Once a month, I run each prompt in the AI products our audience uses, logged out or in a clean profile, from the right country where possible. I run the important prompts twice, because answers vary. And I log everything in a CSV: date, engine, country, prompt, run number, the answer text and the cited URLs.

6:36 Analyze in Python

Then the analysis. In Python, I load the CSV with pandas. For each brand, I write simple patterns, including spelling variants, and flag whether each answer mentions it. Grouping by engine gives the mention rate, the share of answers that name each brand. Summing across brands gives share of voice. Then I split the cited URLs into domains and count them. That most-cited domains list is the gold. It's the publications, directories, review sites and forums that AI answers lean on for our category. It becomes our PR target list. And for accuracy, I score brand answers as correct, partly correct or incorrect.

7:21 Common mistakes and measures

Common mistakes. Treating one screenshot as proof. Chasing tricks like llms dot txt or keyword-stuffed pages, when Google says they get no special treatment. Mass-publishing self-promotional lists on low-quality sites. Ignoring outdated facts in old directories and profiles. And measuring only Google, when your audience uses other assistants too. How do you measure success? Mention rate and share of voice by engine and prompt group. Accuracy of brand descriptions. The domains cited for your category. And Search Console's generative AI impressions for your key pages. Look for change over months, not single runs.

8:01 Recap and try this now

Let's recap. AI search and assistants name and describe brands based largely on what reputable, independent sources say, plus how clearly they can identify you. Google says there are no special optimizations for its AI features, and llms dot txt gets no special treatment. Search Console now shows generative AI impressions, without queries at launch. Measure with a systematic prompt panel across engines, with repeated runs. And improve with coverage, reviews, consistent facts and citable content. Never with manipulation. Try this now. Write twenty customer prompts, run them in three AI products, log answers and cited URLs, and run the script to find your top cited domains. Then list three PR actions.