From Heatmaps to Agent Trace Maps: How Do You See an Agent's Behaviour?

What a heatmap is still good for in the agent era, and what it is not

The value of heatmaps has not changed. For your human visitor they still answer the best questions: which element draws interest, which is ignored, where on the page do they give up? For decisions about CTA placement, form simplification and page hierarchy, the heatmap is indispensable.

What changed is the mix of entities using your page. Traffic from AI agents and agentic browsers grew sharply over the past year. ChatGPT, Claude, Perplexity and Gemini read your page on the user's behalf, compare your product and fill in your form. This visitor is invisible in a heatmap, because it does none of the things a heatmap measures.

• It does not click; it calls the DOM element directly.
• It does not hover; it reads the accessibility tree.
• It does not scroll; it takes the whole page at once, or not at all.

The concept of interest is different for an agent too. A person looks at a heading; an agent parses one. A person notices a price; an agent reads an Offer schema. A person presses a button; an agent interprets an accessible name.

That is why Agent Experience work uses three maps that do not replace the heatmap but sit beside it: an access map showing where the agent reached, a citation map showing what it used as a source, and a task trace map showing at which step of a task it broke off. Together they do for the agent what the heatmap does for the person: they make behaviour visible.

Agent trace map tools and what they do

For a heatmap you install one tool, add the code, and the data flows. For an agent trace there is no single tool; three data sources are combined.

Server logs. The most honest trace of agents lives here. Which URLs user agents such as GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and Google-Extended request and how often, which status code they receive, which files (llms.txt, the sitemap, robots.txt) they fetch. Analytics usually never sees this traffic, because the agent runs no JavaScript and leaves no referrer. Log analysis tools or CDN dashboards make this layer visible.

Visibility tracking. Agents reading you is not enough; they have to use you as a source. Tools such as Brantial run a defined prompt set regularly across the four main engines and record which answers, which page and which context your brand appears in. That answers the question "which page is hottest in the agent's eyes".

Agent task simulation. Defined tasks are run on your site with real agents and every step is logged. It is the agent counterpart of a session recording: it shows which button the agent could not find, which modal it got stuck in, which label it misread.

As a starting point, the free AI Agent Readiness tool gathers the externally verifiable part of all three sources into a single score in 60 seconds. No installation; the domain is enough.

Access maps: the counterpart of the click map

The click map showed which element people touched most. The access map shows which pages agents reach most, and which they cannot reach at all.

Filter your server logs by agent user agent and plot density per URL and a map appears: which pages are hot (fetched often, returning 200), which are cold (never requested), which are red (returning 403, 429 or 5xx).

This map answers questions like these. Do agents concentrate on your product pages or your blog posts? Is your llms.txt being fetched, and if not, does it exist? Which agent is your firewall silently blocking? Is content loaded through JavaScript being seen as an empty page by the agent?

On a click map, "the button nobody clicks" was a design problem. On an access map, "the product category no agent reaches" is an access problem, and it is usually fixed with a single line in robots.txt. Access, the first of Agent Experience's four layers, is measured with this map. Without access the other three layers are meaningless, which is why analysis starts here.

Citation maps: the counterpart of the move map

The move map showed something interesting: the areas a user was interested in but did not click. The cursor rests on a heading, but nothing follows. It revealed the gap between attention and action.

The citation map shows the same gap for an agent: content the agent read but did not use as a source. The access map says a page is fetched often; if the citation map says that page never appears in AI answers, there is a context problem in between.

To build a citation map you define a prompt set about your brand, run those prompts across four engines at regular intervals and map the results per page: which of your pages is cited in which engine, which paragraph is quoted, on which question a competitor beats you.

This map reveals three things.

Quoted paragraphs. Usually short blocks that open with a definition sentence, contain a number and answer a single question directly. Your citation map confirms that pattern for your own site.

Pages read but not cited. Access yes, context no. The heading hierarchy is broken, the definition sentence is missing, schema contradicts the visible copy, or the content is stale. These are the citation map's cold-but-visited regions, and the fastest win lives here.

Wrong citations. The agent cites you but states an old price, a closed branch or a product that is not yours. This is the red region of the map, and it hits brand trust directly.

On the move map, "interest but no click" was a guidance problem. On the citation map, "access but no citation" is a context problem.

Task trace maps: the counterpart of the scroll map

The scroll map showed how far down a page a user went and where they gave up. On long content it was the most valuable map of all.

The task trace map answers the same "where did they give up" question for an agent, but across task steps instead of page depth. When an agent is given the task "find this product, add it to the basket, reach checkout", every step is recorded: which page it requested, which element it found, which tool it called, where it stopped, why it stopped.

Stack a batch of task simulations and a shape very much like a scroll map appears: one hundred per cent at the start of the task, a success curve falling at each step. The step where the curve drops sharply is where the agent gave up.

The most common break-off points are these.

• A cookie wall or full-screen modal the agent cannot close.
• A button with no accessible name; icon only, or just "Continue".
• A fake button built from a <div>; the agent does not know it is clickable.
• A menu that opens only on hover; agents do not hover.
• A form that reports its result only through animation; the agent does not know it succeeded.
• Context lost across a multi-step flow; the basket exists on one page and not the next.

On a scroll map, the finding "users never get this far" led to shortening the content. On a task trace map, the finding "agents break off at step three" leads to fixing a single element at that step. The fix is usually a few lines of HTML, and its effect is measurable.

Improving AX with agent trace data

Read the three maps together and a clear order of priority emerges.

Open access. Fix the pages that show red on the access map and the bots that are blocked. Review robots.txt and firewall rules, move content out from behind JavaScript into raw HTML, publish llms.txt. This step is the equivalent of "the page does not load" on a heatmap; it comes before everything else.

Fix context. Focus on the citation map's read-but-not-cited pages. Put a directly quotable definition sentence in the first 60 to 80 words of every page, repair the heading hierarchy, align schema with the visible copy, add numbers and sources. Fix wrong citations at the source: is the old price left in the schema, or on the page?

Get the task finished. Find the step where the curve drops hardest on the task trace map and fix the element at that step. Turn the fake button into a real <button>, add an accessible name, make the modal closable by keyboard, report the form result in text. Then run the simulation again and watch the curve change.

Cross-read with human data. If the CTA people click most on the heatmap is the same CTA the agent cannot find on the task trace, the problem exists for the machine, not the person: a button that is visually strong and semantically weak. The fix is the code, not the look. The reverse happens too: an element the agent finds perfectly but no person ever clicks is semantically right and visually faint. We covered the button-scale detail of that second case in our CTA article.

What to watch when adjusting the layout

• Think DOM order, not visual position. Moving a CTA up for a person is a CSS job. For an agent, the CTA has to come right after the content it belongs to in the source. Do both.
• Do not break semantics while strengthening an element. Do not turn a button into a <div> to make it stand out. Build the look in CSS and the meaning in HTML.
• When shortening content, move the definition sentence up. If the scroll map says to cut, make sure the quoted paragraph survives; move it into the first screen.
• Same action, same label. Change "Add to basket" to "Buy now" on one page and the language the agent learned breaks. Change a label site-wide or not at all.
• Check the task trace before adding a pop-up. An exit pop-up that lifts human conversion can end an agent's task at step one. At minimum, build one that closes by keyboard and does not obstruct the DOM.
• Re-pull both maps after every change. The heatmap shows the human effect, the task trace the agent effect. If one improves while the other worsens, decide consciously.

Read heat and trace maps together

The heatmap made the human visitor's behaviour visible and for years was the instrument of the principle that data-driven design decisions beat intuitive ones. Agent trace maps carry the same principle to the new visitor.

Click, move and scroll maps say where a person touched, what they were interested in and where they gave up. Access, citation and task trace maps say where an agent reached, what it used as a source and at which step it broke off. A team that reads all six maps knows both users of its page.

Remember: the visitor invisible on a heatmap is not a visitor absent from your site. It merely shows up on a different map. You can start building your own agent trace map with a free agent-readiness scan, and go deeper on the Agent Experience side.