For years, SEOs had a familiar way to track progress. Even when attribution got messy, the basic model held well enough for teams to plan, measure, and defend their work. AI search is different because the value doesn’t always show up as a click or even a ranking.

A brand can be recommended without being cited, and it’s difficult to track conversions when most activities don’t result in direct clicks.

In this AMA, Tom Capper explains why AI search requires a new mental model and what SEOs should actually be tracking, blocking, and optimizing for instead.

1. What is the biggest gap between how SEOs perceive AI and how AI search really works?

The biggest gap is the mental model.

AI didn't suddenly replace lexical search; the latter had been declining long before ChatGPT. Consequently, many SEO assumptions were already obsolete before AI search gained prominence.

It doesn’t mean AI search is completely separate from SEO. When people say “it’s still SEO,” they’re implying some tactics apply across both. That’s fair, but the blend, priority, and way you think about those tactics have to change.

SEOs are still the best-placed people to own this work because we already understand how to optimize for visibility across search environments. 

But we can’t treat AI responses like traditional rankings or assume the old playbook applies without adjustment.

AI search requires SEOs to update the underlying model they use to understand visibility, not just rename the same tactics.

2. How do grounding searches work, and how can brands influence them?

Grounding searches happen when an AI system decides it needs external information to answer a prompt.

For example, ask it to write a poem about butterflies, and it will generate an in-model response using internal training data that could be years out of date. 

Ask it who Tom Capper is and what he's been up to lately, and it will almost certainly run multiple Google searches in the background before answering.

Those background searches are worth understanding for prompt tracking and prompt research. You can inspect them in your browser console, depending on the tool, and Google alludes to them in its fan-out view. 

The searches these tools run can be strange, and I wouldn’t recommend rank-tracking those exact queries. But they show an important shift in how visibility works.

Historically, SEO asked: “How do I get my site to appear for this query?”

Now the question is: “How do I make sure the sources AI systems use describe my brand correctly?”

The system doesn’t care whether the information comes from your website or someone else’s. The grounding search will pull from wherever the best answer lives, including your site and external sources. 

So this is a much greater emphasis on off-site work than we've seen in a long time, or possibly ever, in SEO. In fact, I’ve previously called it digital PR, taking its seat at the adult table.

3. What is the difference between prompt research for AI search and long-tail keyword research for traditional search?

Prompt research is more long-tail than traditional keyword research, which makes exact prompt volume less useful.

You can get prompt volume from sources like clickstream data, but that doesn’t mean it carries the same value as keyword volume.

Most prompts are longer, more conversational, and heavily shaped by context. Higher-volume prompts are often just conversational fillers, like “thank you,” which nobody is realistically going to optimize for, because their meaning depends entirely on the surrounding conversation. 

For traditional SEO, volume helps you decide which queries to prioritize. For AI search, the better approach is to:

• Identify topics of interest
• Track prompts that represent those themes
• Understand who gets cited and how your brand appears
• Monitor how responses change over time around the topic

The query fan-out methodology points in this direction. Rather than asking how many people searched for a specific phrase, it asks what topics and intents cluster around a subject and tracks how the citations change. 

The useful signal is at the topic level, not the individual prompt level, and that’s why prompt research shouldn’t copy the old keyword model. 

4. How should brands think about their bot-blocking strategy?

Brands need to distinguish between training and grounding bots before deciding what to block.

Training bots support the underlying model. In many cases, blocking them may not change what appears in responses for years because large models are retrained or replaced slowly.

Grounding bots work differently. They support live retrieval, web search, and real-time responses to user queries. If you block them, you may affect whether AI systems can access your content when they need current information.

Some publishers, like those mentioned in this study, have valid reasons to block more aggressively, especially when negotiating with AI companies or protecting a traffic-based business model. But for most brands, it makes sense to leave grounding bots open.

This also affects your off-site strategy. Third-party coverage only helps AI visibility when LLMs can access it. If those sites block grounding bots, the coverage may not show up in live responses. So, check if your target sites block grounding crawlers.

Here’s a practical approach to use:

• Know which user agents you’re blocking (training crawlers vs grounding crawlers)
• Avoid blocking grounding bots by default, and if in doubt, don’t block them
• If you decide to block grounding bots, make sure you have a business reason
• Take time to understand how blocking affects both your site and the third-party sites that mention you

5. Is LLMs.txt dead, or is there a legitimate use case being missed?

I was of the view that LLMs.txt was snake oil, but then I found out that Anthropic was requesting it. There’s a curious cycle in AI SEO that we’re not used to in the Google-dominated world — sometimes, one engine’s snake oil can become another engine’s reality.

I can’t speak to how Anthropic is actually using this data. What I can say is that if your AI SEO strategy is Google-focused — which, I believe, in the vast majority of cases it should be — then your LLMs. txt file is doing nothing.

That said, it’s hardly high effort to implement this thing, so you could say, why not?

Limy’s analysis of 515 million LLM bot traffic events found little evidence that major AI bots rely on the file, but it did find agentic use. 

For now, I wouldn’t spend much time on it. There are more useful areas to focus on first, especially understanding which sources AI systems use to describe your brand. But, if you are in a position to stick up an LLMs.txt file in a spare 10 minutes, go ahead.

6. What are the biggest mistakes SEOs are making with prompt tracking?

The biggest mistake with prompt tracking is treating it like rank tracking.

If you look at an AI response and say, “Oh, the first URL shown is mine, therefore I rank first, and this is a success,” you’re measuring the wrong thing.

AI search doesn’t send traffic the same way traditional search did. There’s value, but it’s not the same type of value as traditional search. 

Click-through rates aren't what SEOs are used to, and prompt tracking shouldn’t be judged solely by links, clicks, or URL position.

The better metrics are closer to how you might measure other marketing channels:

• Is your brand visible?
• Are you part of the conversation?
• Are you being recommended?
• Are you associated with the right topic?
• Are authoritative sources describing you accurately?

A recommendation can still be a win even when your site isn’t the one cited.

For example, if someone asks for the best smartphone for their needs and the AI response recommends Samsung and cites PCMag, I wouldn’t call that a failure for Samsung, since they still got the recommendation. Citation is important, but it’s not the only thing that matters.

This is an attribution nightmare, but it reflects how AI visibility works for now. We don’t have the same precise through-attribution we could stitch together ten years ago. Hence, prompt tracking must measure visibility, recommendations, and brand association alongside traditional SEO metrics.

You can easily build a prompt list and track your prompts with the Moz AI Visibility tool. Crystal Carter created a detailed guide that walks you through the process.

7. Does old content get cited less, and will rewriting it fix the problem?

You can answer this for your website, but the answer will vary from site to site.

Start by looking at:

• Citations for the prompts you track in a brand visibility tool
• Log analysis to see which pages AI bots are accessing
• The pages ranking for the grounding queries around your brand and topics

If pages drop in citations, it may be because they no longer rank in grounding queries. Updating this content can help if it leads to reappearing in the searches AI systems use for responses.

However, simply creating new content on the same topic isn't a fix. The priority is identifying which sources AI platforms trust for that subject.

Modern SEO requires thinking beyond your own site. This is the most significant change in the last five years.

While updates matter, a greater opportunity lies in influencing third-party sites that already appear for your brand or category. 

Influencing these external sources often has a greater impact on visibility than rewriting your own articles.

8. Does Google search still exist as we know it in five years?

I expect evolution, not revolution.

A screenshot of a Google SERP will probably look very different in five years, but I don’t think it will look like AI Mode for every query. Google is more likely to move toward a hybrid experience, with SERPs that change more dramatically depending on the search intent.

We already see this today. The results page for a local query, a product query, and an informational query can look completely different, and the trend will likely continue.

I also don’t think Google will stop sending outbound traffic any time soon. It could happen eventually, but it creates a monetization problem because Google’s business model is still built on sending traffic.

Google’s own AI features documentation still frames AI Overviews and AI Mode as experiences that surface links, which supports the idea of a hybrid SERP rather than a single AI answer replacing every result.

Could that change in the future? In theory, yes. 

Agentic e-commerce could replace some of that. But it's unproven, and it doesn't work for every type of business. Five years is a long time, but it's not long enough for Google to rebuild its business model from scratch.

Web Guide is one of the more interesting experiments to watch because it points toward a hybrid future. 

9. What is the most underrated technical move for AI visibility right now?

I think log analysis is one of the most underrated technical moves for AI visibility, especially for larger teams.

Search Console is getting worse, and most AI platforms don’t have a direct Search Console equivalent. It makes log files one of the best places to understand what AI systems are accessing.

In a previous AMA, Jamie Indigo explained that: 

“AI bots can be aggressive and resource-heavy. They will crawl publicly exposed development areas, and security by obscurity does not work. Log files show you what’s actually happening, not what dashboards assume is happening.”

The image below is an example of what it could look like

Log analysis can show:

• Which AI bots are crawling the site
• Which pages they request
• Where they run into access or performance issues
• Whether important pages are being reached at all
• How crawl behavior differs across platforms

It doesn’t replace demand-side tools, because you still need tools like Moz AI Visibility to understand AI presence. 

However, the data you need for AI search is not the same data you used for traditional SEO analysis.

To succeed in the next phase of search, SEOs must rethink how they measure visibility and optimize for these new search surfaces.

Some strategies still overlap, but the way people discover, evaluate, and act on information is changing.