Table of Contents and GEO

A table of contents is a navigation block that lists a page's headings and links to each one. In 2022 its job was to move readers down the page and earn sitelinks in search results. That job has changed. AI search systems do not process a page as a whole; they work at passage level, retrieving and citing individual sections. A table of contents is now a structural signal that declares where those sections begin and end. This article covers what a table of contents actually does in the GEO era, how to build one correctly, and how to measure it.

What a Table of Contents Does in the AI Era

A table of contents is a hierarchical list of the H2 and H3 headings on a page, where each entry links to the matching heading's id. The technical definition has not moved in a decade. What changed is who it speaks to.

In classic search, the payoff was sitelinks: Google parsed the headings and placed four or five jump links under the result. In generative search the payoff comes from somewhere else. ChatGPT, Perplexity and Google AI Mode do not read a full page when answering a question. They locate the section that matches the question, quote it, and attribute the source to that section. A table of contents announces those section boundaries in advance.

That makes it one of the few places where you get a say in how your content is broken apart by a machine, rather than a nice-to-have for readers.

AI Search Retrieves Passages, Not Pages

Generative search runs on retrieval. The system splits indexed text into semantic chunks, selects the chunks that overlap most with the question, and builds an answer from them. The splitting is automatic, and on a loosely structured page it cuts in arbitrary places.

A clear heading hierarchy turns that cut in your favour. When every H2 marks a topic boundary and every paragraph reads on its own, the system takes a coherent unit and the quotation lands whole. When headings are vague and paragraphs depend on each other, the quoted fragment usually arrives incomplete or without context.

A table of contents declares that hierarchy twice: once in the heading tags, once in the link list at the top of the page. The second declaration puts a complete map of the topic in the first screen of HTML, which in practice means the page states what it covers within its first 500 words.

We covered the content-side version of this argument in how to adapt content from featured snippets to AI answers.

How to Build a Table of Contents for GEO

Five steps:

• Give headings stable ids. Use meaningful, permanent values such as id="passage-visibility". Counters like #heading1 carry no meaning and break the moment the heading order changes.

• Place the block after the first paragraph. Above the intro it scatters readers before they have context; at the bottom it serves neither readers nor crawlers.

• Render it server-side. A table of contents generated only by client-side JavaScript is invisible to most crawlers that read raw HTML. The list has to be in the source.

• Stop at two levels. H2 and H3 are enough. A four-level list blurs the hierarchy instead of clarifying it.

• Collapse on mobile, pin on desktop. A collapsible block on small screens and a sticky sidebar column on large ones match how people actually read.

If you use a plugin, apply one test: does the list appear in page source with JavaScript disabled? If not, it contributes nothing to GEO.

Write Headings as Questions, Not Keywords

A table of contents is only as good as the headings it lists. AI systems receive question-shaped prompts, and the closer your headings sit to those prompts, the stronger the match.

"Benefits" answers no question. "What Does a Table of Contents Do in AI Search?" answers one directly. The same applies to "Features" versus "Which Features Improve Your Chance of Being Cited?" A question turns a heading into a matching surface.

The second rule: the first sentence under each heading must answer that heading on its own. The quoted passage is usually that first sentence. Write a warm-up paragraph instead and the system finds nothing clean to lift.

Common Mistakes

• Counter ids. #heading1 and its variants break every link as soon as the order changes, and they mean nothing to a retrieval system.

• Duplicate ids. The browser jumps to the first match and the rest of the links quietly go to the wrong place.

• List text that does not match the headings. When the entry and the H2 differ, both readers and systems lose confidence in the page.

• A table of contents on a three-heading page. Below about six headings it adds nothing and simply fills the first screen.

• Headings used as keyword slots. Repeating one term across seven headings does not improve matching; systems reward coverage of distinct questions, not repetition.

• Expecting emoji. Emoji in headings once showed up in search results. Generative answers do not carry them, so the tactic has no payoff today.

How to Measure the Effect

The effect of a table of contents looks weak in classic metrics, because generative search produces visibility without producing clicks. Measure it in three layers.

First, source level: the generative search report in Search Console shows how often the page appears on AI surfaces. Second, citation level: run your target questions through ChatGPT, Perplexity and Google AI Mode on a fixed schedule and record how many answers cite your brand. Third, passage level: look at where the citation link points. When it arrives with a # anchor, you learn directly which heading is doing the work.

That third layer is the only direct evidence for the investment. We set out how to build the measurement in our piece on zero-click GEO.

The Webtures Approach

We have been building search visibility for enterprise brands for 16 years, and we never treat a table of contents as a standalone tactic. For us the block is the visible face of the heading architecture: we first map which questions the page answers, rewrite the headings against those questions, and only then build the list on top. Reverse the order, dropping in a table of contents and inventing headings to fill it, and nothing moves.

We run this work as part of GEO consulting and track the citation side through citation optimization.

If you want to see which sections of your pages are surfacing in AI answers, get in touch.