AI Assistants Send Visitors to Pages That Don't Exist. That Traffic Is Recoverable.

Most people read Ahrefs' September 2025 hallucinated-link study as a curiosity story: AI assistants occasionally send users to pages that do not exist. True, but incomplete. The more practical reading is that these broken visits are a recovery channel and a product-signal channel at the same time.

Across a 16-million-URL dataset, Ahrefs found AI assistants sent visitors to 404 pages at an average rate of 0.43%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025, versus 0.15%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025 for Google-referred URLs. That is a 2.87xAhrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025 gap.

The immediate objection is fair: AI referral volume is still small. Ahrefs estimates AI assistants deliver roughly 0.25%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025 of the average site's traffic, while Google contributes 39.35%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025. So this is not the first lever a team should pull when traffic is down.

But the low volume is exactly why this is underpriced operationally. Recovering a few hundred or few thousand otherwise lost AI visits is cheap. More importantly, the malformed URLs show what language models expect your site to contain: missing guides, missing comparisons, missing location pages, missing schema-consistent slug families. That is not noise. It is directional demand signal handed to you for free.

Ahrefs ran two complementary tests that are useful to keep separate. In the first test, using Web Analytics, they measured clicked URLs: links users actually visited from assistants. In the second, using Brand Radar, they measured all cited URLs in model outputs, whether or not a user clicked.

The clicked-URL view tells you where user sessions are getting lost right now. The cited- URL view tells you how often models emit dead targets in their answer layer. Both matter, because one affects present traffic and the other predicts future traffic loss as AI use expands.

The interesting middle of the table is Perplexity and Gemini in Test 2. Their cited-URL 404 rates, 0.87%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025 and 0.86%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025, are almost identical to the Google top-20 baseline of 0.84%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025. That pattern is consistent with retrieval systems drawing from search indexes that already contain dead URLs, not necessarily generating those dead URLs from scratch. Copilot sitting below those two at 0.54% in Test 2 fits a similar interpretation through Bing's index.

ChatGPT is the outlier in both tests: 1.01%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025 on clicked URLs and 2.38%Ahrefs, How Often Do AI Assistants Hallucinate Links?, Sep 2025 on all cited URLs. That does not tell you ChatGPT is “bad”; it tells you operating assumptions differ by platform, and a single AI visibility strategy that treats all assistants as one behavior cluster will miss real distribution differences.

Ahrefs points to two causes, and teams should treat both as routine operations work.

This is classic technical debt. A page once existed, got linked, got indexed, got cited, then changed path or disappeared with no 301 in place. Models or search indexes continue surfacing the old path for weeks or months. Users click. Session dies at your 404 page.

Most organizations already have this debt in legacy blogs, renamed service lines, merged location pages, and CMS migrations. AI did not create that debt. It exposed it in a new referral stream where the path is often hidden from marketers because analytics reports are still Google-centric by default.

The second category is more novel. Models generate plausible URLs that never existed but follow your site's historical slug style, such as category nesting or noun-phrase patterns. Ahrefs gives a clean example: `/blog/internal-links/` on ahrefs.com is semantically perfect and structurally believable, but the page does not exist.

This is why dismissing these as random mistakes is operationally expensive. A hallucinated slug often expresses a coherent user need. If many users receive that invented path, you have three choices: redirect it, build it, or deliberately leave it dead and accept wasted demand.

The goal is not to over-engineer a tiny channel. The goal is a lightweight loop that recovers current visits and captures durable content signal while effort stays low.

Build a report where `landing_page` is a 404 and `source / referrer` is AI traffic. Use a rolling 30- to 90-day window. If your analytics cannot classify AI assistants cleanly, start with known referrers and expand from server logs.

Export URL path, sessions, conversions, and first-seen date. You need volume and business impact, not just counts of broken paths.

Check whether each path historically existed. Old real pages should almost always be 301ed to their closest living equivalent. Never-existent pages require a decision: redirect to nearest intent page, or create net-new content if demand persists.

Avoid blanket homepage redirects. They recover less utility and create noisy behavioral data. Intent-aligned redirects are the point.

Fix the top percentile first: URLs with repeated AI sessions, high conversion adjacency, or strategic commercial intent. Teams lose time when they chase one-off weird slugs while leaving recurring failures untouched.

A generic 404 with a search box is weak when users arrive from an AI answer expecting a specific deliverable. Add immediate route options by intent class: service, pricing, comparison, and local pages. Include internal search prefilled from slug tokens where possible.

Aggregate never-existent slugs by stem and topic. Repeated clusters are candidates for new pages, new FAQs, or restructured navigation. This is where the channel shifts from recovery-only to strategy input.

If five variants of a non-existent comparison page keep appearing, you likely have a gap in your public information architecture regardless of whether users arrived from ChatGPT, Copilot, or Perplexity.

Ahrefs reports that 74%Ahrefs, AI content prevalence research, 2025 of newly published webpages now contain some AI-generated content. That matters because hallucinated URLs can propagate: model outputs get copied into posts, posts get crawled, crawled pages become retrieval candidates, and dead paths gain persistence they did not have in one-shot chat interactions.

This feedback loop does not require malice or spam. It only requires routine publishing behavior at scale. As long as AI-written drafts and summaries include outbound references, invented URLs can leak into the broader web graph and keep getting surfaced.

In that environment, maintaining redirect hygiene and monitoring AI referral 404s becomes less like edge-case cleanup and more like baseline channel maintenance, similar to tracking broken internal links or canonical drift.

The operational loop above is straightforward. The harder open questions are distributional: which industries see the highest AI 404 leakage, how fast hallucinated slugs propagate into indexed pages, and whether specific content architectures suppress or amplify fake-path generation across engines.

Those are longitudinal questions, not one-off audit questions. Answering them well requires measuring recommendation and citation behavior repeatedly, by place and by query type, rather than inferring system behavior from one-time snapshots.

Until that kind of measurement is common, most teams can still win here by doing unglamorous basics consistently: track AI-referred 404s, map intent, redirect intelligently, and use repeated hallucinated slugs as evidence of missing content demand.