Scrunch home

A big part of what’s changed is the user experience.

With traditional search, the search engine acts as a gateway. It connects users to webpages where they can retrieve the information they seek.

With AI search, the AI platform is both the gateway and the retriever. It searches, analyzes, and compiles info on the user’s behalf, all in a single interface.

When traditional search was the only game in town, you could measure success on where and how often your site showed up in SERPs.

Not to mention that over 20-plus years of traditional search, the algorithm became less of an enigma thanks to 1.) Google revealing more of what worked (think keyword volume via Search Console/AdWords) and 2.) the market testing and validating what influences search results.

Seeing that we're in the early days of AI search, the honest truth is that there's not a lot of verifiable expertise (as in, confirmed by the AI models themselves) to hang your hat on.

Because that's the case, a sound AI strategy relies on experimentation to tell the difference between correlation and causation. To do that, you need the right AI search metrics in place to first benchmark performance and then measure which growth experiments influence it.

How do I track AI search metrics?

It’s technically possible to get started tracking AI search metrics by hacking together old-school tools and tactics. There are also products built from the ground up to track AI search metrics (we sell one).

Here are your options:

Self-reported attribution

In most cases LLMs mention your brand without linking to your website (this is what’s known as a “zero-click” interaction—when someone gets the info they need without needing to click or fill out a form).

Example: A user prompts an LLM to do research about a product or service relevant to your business. The LLM mentions your brand in its answer but doesn’t link to your site. Later on the user visits your site directly (by typing your URL into their search bar or searching your brand name specifically).

One way to determine if this user learned about your brand via AI search is to ask them when they fill out a form on your site (e.g., sign-up flow, demo request, content download, etc.).

This has its downsides. For one, you can really only ask someone to self-report if they fill out a form or talk to someone at your company. For another, the person could misrepresent how they learned about your brand or simply misremember.

Still, if you’re seeing an uptick in direct traffic or organic traffic from branded search terms, self-reported attribution can give you context for the role AI search is playing.

Referral traffic from LLMs

The obvious way to measure AI search performance is using the tools that are already in place (i.e., tools that track referral traffic sources).

Chances are you’re seeing referral traffic from LLMs like ChatGPT or Perplexity show up in your analytics dashboards. While this is a lagging metric (we’ll get to why in a later section), it’s often the first sign that AI search is a channel worth investing in. The tip of the iceberg, if you will.

For most, the way to uncover LLM referral traffic is via a web analytics tool like Google Analytics (GA4) or HubSpot. Since it’s the most widely used, we’ll cover GA4.

It gets a lot of hate for its UX and modeling system (R.I.P, Universal Analytics), but the standard version of GA4 is free and it can show you referral traffic from AI answers.

Like so:

• Step 1: Open up Google Analytics, navigate to your Traffic acquisition report, and add “Session source” as a filter.
• Step 2: Select the “Matches regex” filter option and copy-paste something like the following (note: Update your inputs based on the sources you want to track):
  (?i)(.*gpt.* | .*chatgpt.* | .*openai.* | .*claude.* | .*gemini.* | .*google.* | .*perplexity.*)

You’ll now have a report that reveals human traffic sent your way from AI platforms and the pages those humans visited.

Once you understand which pages people are landing on from AI search platforms, you can start trying to reverse-engineer the prompts that led them there.

Lots of traffic from AI search hitting a specific product page? It could mean lots of users are prompting LLMs about a certain feature or capability your product has.

The rub: You can’t use GA4 to track zero-click search presence or AI agent traffic, so visibility into AI search performance is murky at best.

AI search products

The methods above are tedious, time-consuming, and imprecise. And, if you’re trying to do AEO/GEO at scale, they’re unmanageable.

Hence AI search products like Scrunch.

We have a pretty strong opinion (*cough* Scrunch *cough*) on which AI search product is best. Admittedly, we’re biased.

Still, we put together a guide to buying products for AI search here.

TLDR takeaway? As you’ll see in the sections ahead, the abilities to self-serve and customize are crucial. You’ll want to be able to add, adjust, and archive to your heart's content as your strategy evolves.

Our advice: Seek out AI search products that provide you with flexibility both in terms of features and your ability to tweak them yourself.

There are four key performance indicators to keep an eye on if you want a complete picture of your AI search performance:

1. Brand presence

Brand presence is how often your brand shows up in LLM answers to prompts relevant to your business.

Within the realm of brand presence, you can dig into share of voice (how often your brand is mentioned versus your competitors) and prompt position (whether your brand falls in the top, middle, or bottom of answers).

The higher your share of voice and prompt position, the more prominently your brand is being spotlighted by LLMs.

This metric helps answer:

• Do I show up in the answers that matter for my category?
• How does my share of voice compare to competitors?
• Which prompts drive the most brand mentions?

2. Citations

Citations are how often your site is directly referenced in LLM answers.

Citations come in the form of links to specific webpages on your site that support the answer the LLM provides (e.g., “According to ACME Corp. research, consumers are…).

You can measure citations both in terms of share of citations across business-relevant prompts and total quantity of citations.

The more your site is cited in prompt responses, the more LLMs view your site as an authoritative and relevant source of information for the prompts you likely care about.

This metric helps answer:

• Is my content being cited by AI models?
• What are the authoritative sources in my category?
• How can I earn or influence citations to improve visibility?

3. Referral traffic

Referral traffic is how many people click through to your website from a link-based citation in an LLM answer.

Keep in mind that most people don’t click—the core reason why users are migrating to AI search is because they get the info they need fast, without ever leaving the AI platform interface. This means that referral traffic from LLMs may look small even though the opportunity for brand discovery, evaluation, and decision-making is massive.

That said, if a person does click through to your site from an LLM answer, they’re essentially the AI search equivalent of a qualified lead. This means AI search referral traffic is less frequent but more valuable.

You can measure this in session counts from AI domains, as well as take it further down the funnel and evaluate it based on conversion rate from AI domains.

This metric helps answer:

• How much human traffic is coming from AI search?
• Which prompts or citations drive the most site visits?
• Do these visits convert better than traditional organic traffic?

4. AI agent traffic

AI agent traffic is how often LLMs send out agents to access the content on your site.

While bot traffic used to be something marketers would dismiss, AI retrieval bots are now your most important visitors. And that’s because these bots are doing the work that a human used to do—that is, they’re crawling your site to conduct research, compare products, and reach a conclusion on behalf of the person behind the prompt.

There are three types of AI agents worth paying attention to, each with their own distinct purpose:

1. Training agents: Scrape websites and transform raw data into structured datasets that sharpen LLM pattern-matching, image recognition, text generation, and so on.

2. Indexing agents: Crawl and analyze web content, building searchable databases for instant retrieval.

3. Retrieval agents: Execute real-time web searches and fetch results to feed LLMs with up-to-date information.

All agents influence how your brand performs in AI search (you can dive deeper into the different types here).

AI agent traffic confirms that the content on your site is accessible to AI models and is either being used in answers today (via retrieval agents) or may be in the future (via training and indexing agents).

You can measure this in terms of agent visit count, agent diversity, and which pages are most popular with agents.

This metric helps answer:

• How often is my site being considered in AI prompts?
• Which AI models are crawling my content?
• Which of my pages do the bots favor?

Should I let AI bots crawl my website?

Speaking of AI bot traffic….

There are some types of businesses—specifically those dependent on content monetization—where it doesn’t really make sense to hand over content to LLMs.

Think publishers like The New York Times or content creators on YouTube. Organic search traffic isn’t part of the customer journey for them—it’s the entire business model.

Web infrastructure companies like Cloudflare are defaulting to blocking AI web crawlers in part to help these kinds of businesses.

Even so, unless the content on your site is monetized, we recommend that you allow AI agents to crawl it.

The more that LLMs know about your brand, products, and services, the likelier it is that they’ll represent them accurately in responses.

Why? Because in most AI search products, this information is used to generate prompts, make content recommendations, and more.

The more thorough you are in pre-planning (e.g., thoughtful persona descriptions, comprehensive competitor set, etc.), the better your results will be on the other end and the more time you’ll save down the line.

🔌 Shameless plug: Scrunch makes this easy by using AI behind the scenes to identify personas, competitors, topics, etc. based on your brand’s website domain. That said, this info should be used as a starting point to take some of the heavy lifting away. Review and adjust based on your brand knowledge.

Brand context

First and foremost, what’s your brand’s name?

Seems simple, but it can get complicated if you go by multiple names or are a sprawling brand like Disney (e.g., Disney, Disneyland, Disney World, Disney+, Disney Store, Disney Cruise Line, etc.). Include all applicable names and websites.

You’ll also want a succinct description of the products and services you’re known for (or want to be known for). Think the kind of boilerplate copy that usually lives in the footer of a website, but more explanatory and with less marketing jargon.

Location

How granular you get depends on where your customers are and how important locality is to your business.

For lots of companies something like “North America” or “EMEA” will do the trick. Others may want to get specific about cities and zip codes depending on what they do and who their target audience is.

Precise locations are especially important for regional brands (think car dealerships) trying to gin up interest from local customers.

Industry

The industry or market your business operates in or addresses is especially important for benchmarking and competitive analysis.

This info is used for prompt volume approximation.

Example: The automotive services industry sees 10 million prompts per day.

Customer personas

Think job titles, business segments, pain points, and priorities.

Ideally this info is copy-pasted directly from your product marketing team.

Again, you may want to get granular here on location (e.g., BMW owners and buyers in Orange County, CA).

Competitors

Who’s share of voice are you hoping to eat into?

This info is helpful for benchmarking how your brand stacks up with others in the same space.

Similar to your own brand, you’ll want to account for alternative names or websites (e.g., Meta could include Facebook, Instagram, Threads, WhatsApp, Oculus, etc.).

Topics

Certain topics will matter to your brand (e.g., if you sell an HR platform, topics like hiring, workforce management, and payroll will probably be top of mind).

Prompts are a subset of topics.

You can theoretically track as many topics as you want, but figure out which ones are a must-win for your brand.

AI search monitoring tips

A few tips to help you get started:

Pick your primary goal

Before you start loading prompts and tracking trends, it’s helpful to think through the context above and home in on your primary objective (beyond world domination).

Is the goal to increase visibility from a specific persona? Gain share of voice versus a specific competitor? Build out a citation strategy? Improve brand sentiment on AI prompt responses?

This will influence the analytics you should prioritize and the actions you should take.

Tap into your PMM team

It may look different at your company, but in many instances, the product marketing team will have all the context you need for AI search monitoring—product positioning, key personas, top competitors, etc.—already documented somewhere.

This is like a cheat code for feeding your AI search product high-quality intel.

Save yourself some time and effort by working with your PMM team to dig it up.

Find your competitive sweet spot

Every brand will have a Goldilocks zone of competitors.

Depending on how crowded your market is, you’ll generally want to aim for tracking performance against 5-10 competitors.

It’s also worth being honest with yourself about who you’re really competing with. You should be in the same echelon as the competitors you track.

Put your SEO planning to work

There’s a good chance you already have a list of topics or keywords you use for SEO tracking.

This is a great starting point for selecting AI search topics.

Don’t have a spreadsheet handy and not using Scrunch (we do this automatically for our customers)? Feed the context above into your LLM of choice and ask it for recommendations.

This way, whenever you need to track a new prompt (or a whole batch of prompts for a campaign), you can just slot it right in.

More importantly, you’ll be able to ask insightful questions and get trustable answers back more quickly, like:

• Which high-intent prompts are we cited for alongside [insert competitor]?
• What is our brand presence at the [insert persona] level?
• What was our reach across LLMs for [insert campaign]?

You can grab an actionable Google Sheet template of our framework here. It follows this structure:

Prompt clusters

These are your high-level containers. Think of them like keyword groupings in SEO—prompts that share a common theme across the dimensions below.

Examples include:

• Competitors and alternatives
• Product use cases
• Outcomes and ROI
• Campaigns

Cluster your prompts in whatever way makes sense for your brand.

Funnel stages

Prompts within each cluster should be mapped to buying funnel stages. At Scrunch we use the following four stages:

1. Awareness: Pains, problems, friction
2. Advice: Guidance, frameworks, how to think about “X”
3. Evaluation: Exploring solutions, features, capabilities
4. Comparison: Choosing between options

This makes it easier to make sure you have prompt tracking coverage across the entire funnel, as well as map follow-up prompts to their corresponding stages.

Personas

Some prompts will be particular to a persona within your ideal customer profile.

When that’s the case, it’s a good idea to append a persona to them (e.g., C-level, operations, engineering, etc.).

This makes it easier to isolate prompts for specific target audiences.

Tags

Tags let you append custom metadata to prompts.

This is especially helpful for keeping track of time-bound prompts (think brand campaigns) and bespoke designations (think a specific product integration or partner).

It’s also just a good practice to create connective tissue across topics (think using “pricing” as a tag).

Regions

If you’re operating at a global scale, it’s also worth tracking prompts by region.

Different geographic locations may have their own nuances (e.g., how a question is asked, regulatory compliance considerations, etc.).

Putting it all together

Thinking through this framework upfront saves a lot of time and future tech-debt headache.

To test if it will scale, run a thought exercise of adding a new topic cluster to the mix (e.g., a new product line or a new use case).

It should feel easy to:

• Generate or ideate prompts that fit within said topic cluster
• Layer into your existing AI search structure without being disruptive
• Isolate to track performance of the new topic (or any new dimensions within it)

If that’s the case, you’re in good shape and ready to start generating prompts to monitor.

How do I source prompts?

Prompts—as in the actual words human beings use to interact with LLMs—are what you need to track to understand metrics like brand mentions and citations, as well how those metrics trend over time.

Let’s start with how to select which prompts to track before loading them into your AI search product.

There are a few options to get you up and running fast:

Use our Prompt Generator

We built a free Prompt Generator.

Just drop in your domain and it’ll generate prompts you should be paying attention to and show if your brand’s showing up in AI search.

Borrow from SEO

Like we mentioned above, if you already have an established list of topics or keywords you’re tracking for SEO purposes, it’s a solid starting point.

The catch is that the typical prompt is a lot longer than the typical search query.

A tool like ChatGPT or Claude can help you convert a list of topics or keywords into prompts. Literally load up your list and ask the LLM to give you its best guess at corresponding prompts in return.

🔌 Another shameless plug: Scrunch does this, too. Either copy-paste your keywords or import them into Scrunch using a CSV and we’ll convert them into prompts for you.

Borrow from search ads

Your organic rankings and Search Console tell you what’s working on the earned front (i.e., the prompts that your content is already optimized for).

The other aspirational side of the coin: paid search terms.

These are the terms you’re not ranking for but are paying to show up for. In AI search, they’ll convert to the prompts you’ll want to create content for.

To turn your paid terms into prompts, locate them in your ad platform (or SEO keyword tracking software) and export.

In the Google Ads platform, it’s not easy to find. From the home screen, you’ll want to locate Audiences, keywords, and content.

Then select: Keywords → Reports → Performance summary → Search keyword.

Then (and this is an important step, as it gives you the real keywords used), add a column for “Search term” to show “actual searches that a significant number of people have used, and that resulted in your ad being shown.”

Export that and convert it to prompts the same way you’d go about it above for organic terms.

Borrow from customers

Two sources to mine for prompts that are pure gold:

1. Call transcripts
2. Support tickets

Call transcripts give you the objections and jobs to be done that come up on sales calls. If these questions are being asked on calls, chances are they’re being asked by others who are searching for a product like yours.

Mining prompts from this source can be a challenge. What you want to do is extract as many relevant transcripts as possible and feed them into an LLM for analysis. Ask it to draw out common patterns and use cases that come up over and over again. That should be a great starting point for prompts to start tracking.

Further down the funnel are support questions that customers are asking. These are questions users of your product ask an AI chatbot (e.g., we use Fin from Intercom) or that are submitted to your support team.

Most support platforms will allow you to extract a log of support questions. With that export, you can load questions into an LLM for analysis and pattern match to find the questions that are asked repeatedly.

From there you can use them as prompts to monitor and optimize for.

Bonus: When you create content for these prompts, LLMs become an accurate source of support (better than the user taking a hallucination at face value, at the very least). Not to mention your support team (or AI agent) will thank you.

Borrow from Reddit, Quora, and other communities

Not only do LLMs cite these sources often, they’re also rich with questions the market is asking about product use cases, the category you’re in, and sometimes your product specifically.

The best way to extract prompts from these sites is to be in them yourself. That is, immerse yourself in a Subreddit, identify when a relevant question is asked, track it, and optimize for it.

Next best option is to use a tool or prompt an LLM to do that research on your behalf.

Prompt for prompts

If you’re starting from a blank slate, you might as well ask AI for an assist.

Pro tip: Don’t try to one-shot it. Chunk out your prompt generation at approximately 20 prompts at a time.

This will help you get more specific prompts for different topics, personas, stages of the funnel, etc. Just be sure to add in the context necessary to generate prompts that will matter to your brand.

Write prompts out by hand

Tedious? Yes. But this option offers maximum control.

If you’re going to come up with prompts the old-fashioned way and add them to an AI search product, just don’t do it one at a time.

Write them all out in a list first and then copy-paste the full list into the product (and, of course, make sure you’re using a product that allows for custom prompt additions and editing).

How do I track and analyze prompts?

Now for the fun part.

You’re ready to start actually tracking prompts so that you can benchmark your brand performance and improve it over time.

Before we dig deeper, it’s worth pointing out that you can technically do this manually across AI platforms versus investing in an AI search product.

You’d likely want to build visualizations on top of the data you uncover to make it more human-readable, as well as implement a number of controls to make sure the data is trustworthy. But it is doable.

It looks like this:

Log in to an AI platform → Run your prompts → Log your results in a spreadsheet or database → Repeat across every AI platform separately → Repeat again and again at regular cadences to monitor trends.

If that sounds like a soul-destroying timesuck to you, we agree.

For that reason, we’ll assume you plan to use an AI search product like Scrunch.

Data analysis

Start with the raw numbers.

There are a few key metrics any AI search product worth its salt should be able to show you:

• Presence: How often your brand shows up in a prompt result
• Position: Where in the prompt result your brand shows up (i.e., top, middle, bottom)
• Citations: How often your website is referenced as a source in prompt results
• Sentiment: The attitude toward your brand in the prompt response (i.e., positive, neutral, or negative)

Keep in mind that sentiment will probably look overly positive, as LLMs (maybe with the exception of Grok) tend to adopt a rosy tone in responses.

Data visualization

Next, use visualization to tell a story with the numbers.

AI search products should represent the metrics mentioned in the section above as a change over time—a bar chart comparison, a trending line chart, or something similar.

This makes it easier to spot trends and anomalies, run competitive comparisons, and so on.

Reporting levels

Most AI search products will offer different levels of granularity in their reporting (how granular depends on the product):

High-level summary

This is the kind of data you’d share with a CEO or board.

• Answers: How your brand’s AI search performance looks on the whole
• Intended audience: C-suite executives who want a bird’s eye view
• Example: Chart showing trends across all business-relevant prompts for a given time period

Detail drill-down

This is the kind of data (filtered using the prompt tracking framework above) you’d use to start informing your AI search strategy.

• Answers: How your brand’s AI search performance looks across different platforms and stages of the buyer's journey—and how it’s responding to specific initiatives
• Intended audience: Marketing leaders who want more details, SEO and growth practitioners who want to report on their KPIs
• Example: Chart showing comparison to competitors for prompts associated with a specific topic

Prompt-level drill-down

This is the kind of data you’d use to get in the weeds for specific prompts to identify trends and uncover insights you can use to improve brand presence or citations within responses.

• Answers: How your brand’s AI search performance looks at the prompt level, including presence, citations, and share of voice versus competitors
• Intended audience: Individual contributors who want to diagnose challenges and opportunities for must-win prompts
• Example: Chart showing comparison to competitors for a specific prompt in a specific LLM based on brand presence and citations

Model-level drill-down

This is the kind of data you’d use to understand how your brand is winning or losing across specific AI platforms.

It would give you all the same data as the prompt-level drill-down reporting above, but at the model level (e.g., ChatGPT, Claude, Perplexity, etc.) via daily snapshots across presence, position, citations, and sentiment.

Exports and API usage

It’s common for data teams to want to extend AI search data to their data warehouses or business intelligence tools.

This requires the ability to export your data. Even better if you can use an API to directly connect data to a system of record.

🔌 Another shameless plug: Scrunch lets you do both. You can export any prompt data at any level to a CSV or use our API to sync data with your system of choice (see more info here).

As a starting point for tracking agent traffic, you’ll want to be able to answer the following:

How much AI agent traffic is my site getting?

Knowing how much AI agent traffic your website gets helps you understand if you’re being referenced in AI search responses and how often.

If your AI search product has an agent traffic monitoring feature, you should be able to connect your site to it to surface this data (here are step-by-step directions for how to do it with Scrunch via Akamai, Cloudflare (options one and two), Vercel, and WordPress).

This should show you:

• Total traffic over a set time period
• Traffic distribution between training, indexing, and retrieval agents
• Comparison between the current period and the last period (%)
• Top agents and when they were last seen
• Top webpages accessed by LLM bots
• Recent bot requests

Which AI agents are visiting my site most frequently?

Knowing which AI agents land on your site most often helps you understand which platforms your brand is performing best on—and which ones could use some love.

Your AI search product should show you how much agent traffic you’re getting from different AI platforms—both in real time and in aggregate—over a given timeframe, as well as how that traffic is increasing or decreasing over time.

Which of my site’s webpages are most popular with AI platforms?

Knowing which webpages get the most attention from AI platforms helps you understand what your audience is searching for and what kind of content LLMs find most helpful to answer them.

Your AI search product should show you which pages on your site are being visited most frequently by AI agents and the number of visits to those pages over a specific time span.

Which AI agent visits indicate commercial intent?

Knowing which types of agents are visiting your site helps you understand whether your content is being used for training, indexing, or retrieval purposes.

Your AI search product should show you which visits are related to each of those buckets, as well as the total count for each over a given time period.

How do I track AI search volume?

Prompt performance is the lifeblood of any AI search strategy.

That said, if you’re trying to optimize for AI search, you also want to know what people are asking AI about in the first place—and how often.

Here are a few examples:

Only AI platforms themselves have total visibility into AI search trends, which means AI search products must turn to panel data (same as how SEO tools measure monthly search volume for keywords).

That means this data is directionally useful, but not 100% precise.

With that in mind, make sure you’re working with a trustworthy AI search product that’s transparent about where and how it’s gathering this data (for instance, at Scrunch, we’re only offering AI search trend data at the topic level to start because there’s simply not enough trustable prompt-level data).

Tracking AI search volume will help you understand:

How popular are different topics in AI search?

If your product provides AI search trends, it should allow you to select topics relevant to your industry (and suggest topics of its own) and view conversation volume (i.e., the number of related prompts) over a customizable timeframe.

This will give you an idea of what people are asking AI about and where you may have opportunities to get in front of more eyeballs.

How does my brand perform in AI search for popular topics?

Just as important as which topics are most popular is how often your brand shows up.

Your AI search product should display your brand presence across all the AI platforms you care about, as well as what the general sentiment expressed toward your brand is in AI responses.

This will show you how you’re performing for business-relevant topics as a whole.

How do my competitors perform in AI search for popular topics?

It’s helpful to know how other brands are showing up for certain topics, not just your own.

Your AI search product should give you a breakdown of the top brands for different topics as well as tell you how your brand presence compares to theirs over a set time span.

This will show you which competitors have the highest share of voice for specific topics.

What are people actually asking AI about different topics?

It’s important to understand what topical AI conversations look like at the prompt level (i.e., what people are actually asking and the real responses).

Your AI search product should provide real examples of prompts and responses for different topics.

This will show you the type of language people are using, provide better context for the “why” behind the prompts, and make it clearer what LLMs are saying (or not saying) about your brand.

That’s it for chapter 2. In chapter 3, we’ll dig into what you can learn from AI search monitoring and how to take action on it to improve performance.