You have a page you are proud of. It is linked from your pillar post, your nav, and three related articles. Then you ask ChatGPT the exact question that page answers, and it cites someone else.
That stings. It also raises a fair question: do internal links help AI citations at all, or is linking just an old SEO habit we carried into a new channel?
Here is the honest answer, and it is a useful one. Internal links do real work, but not the work most people think. They improve the conditions under which your page can be found and understood. They do not, on their own, decide whether an AI engine cites you.
That gap between "helps" and "decides" is where most of the confusion lives. Let's clear it up. By the end you will know exactly which part of the pipeline internal links touch, which part they do not, and how to stop guessing about the difference.
The short answer: internal links help you get found, not chosen
Lead with this and you will not go wrong: a crawlable internal link can help a search crawler discover your page. Discovery is not indexing. Indexing is not retrieval. Retrieval is not citation.
Four different steps. Internal links show up strongly in the first one, plausibly in the second, and indirectly after that.
Google is direct about the first step. It describes URL discovery as an ongoing process, because there is no central registry of every page on the web. Google finds a new URL by extracting a link from a page it already knows, like a hub page linking to a new article. Sitemaps are another route in.
So the link is a path. A real one. A page with no route in is harder to find, though a sitemap or an outside mention can still surface it.
What a link is not is a promise. Google says plainly that even a page following all of its requirements is not guaranteed to be crawled, indexed, or served. Every step after discovery has its own gate.
That is why "do internal links help AI citations" is the wrong shape of question. The better one is: which gate am I actually stuck at?
What has to happen before a page can be cited at all
Picture the sequence as a hallway with five doors. Your page has to get through all of them.
- Discovery. A crawler learns the URL exists, often from a link on a page it already knows.
- Crawling. The crawler fetches the page. Google says it does not crawl every URL it discovers. A login wall, a server error, or a robots.txt rule can stop it here.
- Indexing. The system analyzes the page and decides whether to store it. Not guaranteed. If there are duplicates, a canonical gets picked and the others sit out.
- Retrieval. For a specific question, the page enters the candidate set of possible sources.
- Citation. The engine picks it, links it, and maybe leans on it while writing the answer.
Internal links stack the deck at door one. They help a little at door three, because context helps a system understand what it is storing. After that, their influence gets thin and indirect. That is the honest shape of the internal links AI search story, and it is narrower than the advice you have probably read.
This is worth internalizing, because it changes what you do when a page underperforms. If nobody can reach the page, linking fixes something real. If the page is indexed and still never cited, adding a sixth internal link is not the fix. The problem is behind a different door.
One more thing about door one. Google says it can generally crawl a link when it is an HTML anchor element with an href attribute, and that links in other formats often are not parsed. A link built out of a click handler with no real anchor may not create the path you think it does. That is a plumbing detail, not a ranking secret, and it is worth checking once.
Different engines guard these doors differently, too. OpenAI names OAI-SearchBot as the crawler that surfaces sites in ChatGPT search, and says sites that opt out will not be shown in ChatGPT search answers, though they may still appear as navigational links. That access is separate from GPTBot, which is about training data. Perplexity says it respects robots.txt, and that if you disallow PerplexityBot it will not index your full or partial text, though it may still hold your domain, headline, and a short factual summary.
Do not take one engine's behavior and assume the rest match. They do not.
Link text does real work, just not the work you expect
Here is the part people skip, and it may be the most useful thing in this article.
Google defines anchor text as the visible text of a link, and says it tells both people and Google something about the page being linked to. Its advice is to make anchor text descriptive, reasonably concise, and relevant to both the page holding the link and the page it points at.
Google also says the words before and after a link matter. The whole sentence gives context. It warns against chaining links together, because that leaves each one with no surrounding text to explain it, and it warns against stuffing every related keyword into anchors, which it treats as a spam violation.
Read that again, because it flips a common instinct. The value is not in matching your target keyword exactly. The value is in a link that makes sense in its sentence and sets an accurate expectation for where it goes.
Google's own internal-linking guidance adds two plain recommendations: give every page you care about a link from at least one other page, and link related resources in context.
So does this context carry over to answer engines? Partly, and honestly. Google says link text helps it interpret the destination. Google says internal links help it make sense of a site. Answer engines often lean on search systems and indexes before they generate anything. Put those together and coherent internal context plausibly helps a page be found and understood in those systems.
What nobody has shown is a disclosed, universal rule that ChatGPT, Perplexity, Gemini, or Claude boosts a page because of its internal-link count, its click depth, or its anchor wording. Treat "topical proximity is a direct citation signal" as a reasonable hunch, not a fact. Internal linking LLM ranking claims deserve that same care until someone names the engine, the variable, and the result.
What Google actually says about its own AI answers
If you want one source that gets closest to the question, it is Google's guidance on AI features, and it is refreshingly boring.
Google says existing SEO best practices remain relevant to AI Overviews and AI Mode, and that there are no additional requirements or special optimizations needed. For a page to be eligible as a supporting link, it has to be indexed and eligible to appear in Google Search with a snippet. And in the foundational practices, Google explicitly lists making content easily findable through internal links.
That is the strongest documented support you will find. It is real. Notice what it says and what it does not.
It says internal links are foundational for findability. It does not say an internal link earns a citation.
Google also describes query fan-out in AI Mode and AI Overviews: the system may issue multiple related searches across subtopics and data sources to build a response, and while doing that it identifies more supporting pages and can show a wider set of links than classic search.
Careful here, because this is where a myth is born. Fan-out means more searches, not a crawler walking your site's link graph the way a person clicks around. Your page can get found through a subtopic question rather than the user's original wording. That is a good reason to cover a topic properly. It is not evidence that the engine traverses your menu.
ChatGPT works differently again. OpenAI says ChatGPT searches the web when a question might benefit from it, sometimes partners with other search providers, and typically rewrites your question into one or more targeted queries. It says ranking is based on a number of factors meant to surface reliable, relevant information, and that there is no way to guarantee top placement. The factors are not published.
That is the whole picture. The internal links AI search crawlers can follow may help your page reach an upstream index. What happens after that is engine-specific and mostly undisclosed.
What the research tested, and what it did not
This is where a lot of confident advice gets built on the wrong foundation. Let's look at what the studies actually did.
The original 2024 GEO paper tested nine methods for improving visibility in generative-engine responses against an unmodified baseline. It reported that the best methods improved the baseline by 22% on Position-Adjusted Word Count and 37% on Subjective Impression, and could boost source visibility by up to 40%. It even tested against Perplexity in the real world.
Impressive. Also worth reading closely. Those methods were textual: adding relevant citations, quotations, and statistics, plus changes to fluency, authority, and readability. The study did not test internal-link architecture, link counts, anchor text, graph distance, or site structure. If you see that 40% number attached to a linking tactic, it has been borrowed from an experiment that never ran.
A 2026 paper called FeatGEO went deeper on features. It treats pages as combinations of structural, content, and linguistic properties, things like headings, length, list density, introductory summaries, unique information, quotations, and cited sources. It evaluated on GEO-Bench, 10,000 queries across 25 domains and nine sources, with GPT-4o-mini, Gemini-2.5-flash, and Qwen-plus generating answers. It found that high-level discourse organization and information structure matter more than surface word choices.
Useful. But its setup injects the candidate page alongside the top five already-retrieved pages. The page is assumed to be in the room already. So the finding is conditional on retrieval, and again, no internal-link or site-graph experiments.
A third 2026 framework splits generative visibility into two stages that we tend to smash together: citation selection, where the platform triggers search and picks a source pool, and citation absorption, where a selected source actually contributes language, evidence, or structure to the answer. Across 602 controlled prompts on ChatGPT, Google AI Overview or Gemini, and Perplexity, it logged 21,143 valid search-layer citations, 23,745 citation-level feature records, 18,151 successfully fetched pages, and 72 extracted features. It found ChatGPT cites fewer sources but with higher mean influence, while Perplexity and Google cite broader sets with lower per-source absorption. Pages with high influence tended to be longer, more modular, and more semantically aligned with the answer.
That distinction is genuinely helpful. Being cited and being used are not the same thing. And once again, the framework reports no analysis of internal links, anchor text, navigation, site structure, or graph distance, and it cautions that its features do not causally force citation.
So here is the fair summary of the evidence base. Crawlable internal links can improve discovery and help systems interpret how pages relate. Classic ranking and AI citation run on different pipelines and should be measured separately. And the published experiments have mostly tested page content and structure after retrieval, not links.
That is less than the internet promises you. It is also solid ground to stand on.
Where the internal-link claim gets oversold
Six claims to retire, because none of them survive the evidence:
- "Add five internal links and the page will get cited."
- "Internal links are a direct ranking factor in every LLM."
- "Topic clusters guarantee citations."
- "If a page ranks in Google, ChatGPT will cite it."
- "More internal links are always better."
- "Exact-match anchor text is an AEO tactic."
None of that is supported. And the honest version is easier to work with anyway.
Internal links are infrastructure. The site structure AI answers depend on is a floor, not a ladder. They cannot rescue a page that does not answer the question. They cannot un-block a page that robots.txt is keeping out. They cannot get an unindexed page into a candidate set, make thin evidence convincing, or beat a source that simply fits the query better.
They also cannot force a citation, because an engine may cite only part of an answer, choose a different source, or use your page without visibly crediting it.
If that feels deflating, flip it around. It means the effort you were about to spend on a linking sprint is probably better spent on the page itself. That is good news. It is cheaper and more in your control.
What to measure instead of assuming
The way out of this whole debate is measurement. You do not have to settle whether internal links cause citations in the abstract. You can just watch what your own pages do.
Start by keeping five things separate in your head, because collapsing them is what creates false conclusions:
- Indexed. Is the page stored and eligible to appear at all?
- Mentioned. Does an AI answer name your brand without linking you?
- Cited. Does the answer attach a visible source link to your page?
- Which URL. Which specific pages earn those citations, not just the domain?
- Who else. Which competitor pages get chosen for the same questions?
Vocabulary is doing quiet work here. Mentioned, retrieved, ranked, cited, and used to write the answer are five different events. Treat them as synonyms and every conclusion you draw will be a little bit wrong.
Also resist reading a citation count as a score. Bing's AI Performance report shows how often a site's content is referenced in supported AI experiences and which URLs are visibly cited, and Bing warns explicitly that the number reflects citation frequency, not importance, ranking, or the page's role in the answer. A source can be cited and barely used. Another can shape most of the answer.
This is the part where tracking beats theorizing, and it is why AI visibility tracking exists as a category at all. DeepSmith tracks mention rate, citation rate, and share of voice per prompt, shows which of your pages AI actually cites and which prompts drive them, and maps your site's topics next to your competitors' so you can see where coverage is thin. That does not settle the causal question about links. It does tell you which pages are getting chosen, which is the question you actually need answered this quarter.
So, does internal linking affect AEO? Yes, at the access and interpretation layers, where it is documented. Beyond that, stop assuming and start watching. That is the part of your site structure AI answers really turn on: whether the page is reachable, understandable, and genuinely the best answer available.
You are probably closer than you think. Most sites that struggle here are not badly linked. They are unmeasured.
Pick your ten most important pages this week. Confirm each one is reachable from at least one other page, in a normal sentence, with link text that says where it goes. Then set up tracking for the questions your buyers actually ask, and let the data tell you what to fix next. If you want that tracking and the content production in one place, you can start a free DeepSmith trial and see your real citation data before you commit to anything.
