How to Mitigate the Impact of AI Assistant Hallucinated Links on Your Brand's Traffic

What changed in AI hallucinated links?

The change is that AI referrals now need their own URL-quality review. In a study published on September 2, 2025, Ahrefs examined 16 million unique URLs cited by ChatGPT, Perplexity, Copilot, Gemini, Claude, and Mistral. It found that AI assistants sent visitors to likely 404 pages at an average rate of 0.43%, while Google Search referrals in the same analysis had a 0.15% 404 rate. That makes the AI-assistant average 2.87 times the Google referral baseline.

This does not prove every failed AI-referred URL was invented. Ahrefs explicitly notes that a 404 can result from a deleted page, a moved page without a redirect, or a URL that follows a plausible but nonexistent pattern. The distinction matters. A moved page may have a strong replacement. An invented path may have no honest equivalent at all.

The more useful operational change is simple: do not treat AI traffic as one referral bucket. Measure the landing URL, HTTP result, source assistant, and user outcome. This turns a frustrating dead end into a review queue that can be acted on. It also lets a brand see whether an assistant is naming the company but sending the resulting visitor somewhere unusable.

The issue belongs inside AI search visibility work, not only technical SEO. A citation that reaches a broken URL can fail the person who clicked it even when the brand appeared in the answer. That is why traffic quality should sit beside Citation Share, Answer Presence, and competitor monitoring. Rankings got you found. Citations get you chosen. A broken cited link can interrupt the second half of that journey.

Ahrefs also ran a cited-URL test, rather than a clicked-URL test. Among URLs available in its crawler database, ChatGPT had a 2.38% 404 rate, compared with 0.84% for URLs from Google's top 20 results. Perplexity and Gemini were closer to that Google benchmark at 0.87% and 0.86%. These are study results, not a universal rate for every website, but they are enough to justify a regular audit.

1. Create an AI-referral view in analytics before judging the scale of the problem.
2. Record the exact landing path and HTTP status instead of relying on a generic not-found page report.
3. Separate a former real URL from a path that has no publication history.

Who does AI link hallucination affect?

AI hallucinated links affect any brand that can be mentioned by an answer engine, especially a site with a large content archive, changing product URLs, or a predictable URL structure. A visitor may receive a citation to a page that has moved, a page that was removed, or an imagined page that resembles a normal site path. The visible brand mention does not guarantee the click reaches useful information.

B2B software companies are exposed when AI assistants answer category questions, product comparisons, implementation questions, or pricing-adjacent questions. A fabricated documentation path can send a high-intent visitor to a dead end. Local and multi-location businesses face a related issue when an assistant creates location or service paths that look plausible but do not exist. The result can be a lost booking path rather than a lost pageview alone.

Content teams are affected too. A hallucinated URL sometimes exposes weak redirect hygiene. If a site has changed its information architecture, retired a guide, or consolidated a product page, an old cited URL can be a legitimate destination that needs a relevant replacement. Google's guidance says redirects tell visitors and Google Search that a page has a new location, and says permanent server-side redirects are the best option when a URL has permanently moved to a new location.

Not every missing path deserves a redirect. Redirecting a made-up URL to the homepage or a loosely related page may hide the evidence and frustrate the visitor. The correct standard is relevance: redirect a former URL when there is a clear successor that answers the same need. Leave a true non-page as a 404 or a helpful not-found experience, while preserving the evidence in the audit log.

This is also an answer engine measurement problem. The assistant, the cited URL, the page state, and the final conversion path need to be evaluated together. A brand that only counts mentions can miss the difference between being named and being reachable.

1. B2B SaaS teams should inspect product, comparison, documentation, and integration paths.
2. Multi-location businesses should inspect service and location combinations that look easy for an assistant to construct.
3. Publishers should inspect retired URLs, changed slugs, and common archive patterns.

How should you find AI hallucinated links on your site?

Start with your own analytics, because it can show the real landing URLs that received AI-referred visits. Build a report that filters sessions by AI-assistant referrer, then group the results by landing page. Add the HTTP status or a not-found marker from server logs, a monitoring platform, or a page-title convention. Ahrefs used page titles containing 404 or not found as a proxy in its clicked-URL analysis, while noting that this can undercount failed pages.

Do not stop at a total 404 count. A total tells you there is a problem, but not whether it comes from ChatGPT, Claude, Perplexity, Gemini, Copilot, or another source. Break the report down by referrer where the data supports it. Then capture the full requested path, the first known date, sessions, engagement, conversions where available, and whether the URL once existed.

Next, classify each path. Search your CMS history, sitemap archives, redirect map, analytics landing-page history, and server records. The purpose is not to force every URL into a fix. The purpose is to decide whether the path is an old real page, a typo-like variation, a generated combination, or a URL with no evidence of ever existing. Keep the original request intact in the record. It is the evidence that lets the team identify repeated patterns later.

A small recurring audit is more useful than a one-time cleanup. Review the highest-impact AI-referred 404s first, then compare the new set with previous runs. This supports a clearer view of traffic health around citations. It can also reveal a content gap when people repeatedly arrive on a credible topic path that the site does not cover, although a content decision should still be based on demand and editorial fit rather than one invented URL.

If your analytics does not reliably identify the assistant, document that limitation rather than guessing. Referrer behavior can change, and some visits may not carry a useful referrer. A clean audit distinguishes what was observed from what could not be measured. That is more credible than treating every unexplained 404 as an AI hallucination.

1. Filter analytics for known AI-assistant referrers.
2. Export landing paths with sessions and HTTP outcome.
3. Classify each path using site history before changing redirect behavior.
4. Prioritize URLs with meaningful traffic or conversion intent.
5. Keep a dated log so repeated patterns are visible.

What should you redirect, and what should remain a 404?

Redirect only an old real URL that has a clear, relevant successor. Google explains that a permanent redirect signals that the target should be canonical, and recommends a permanent server-side redirect when a page has permanently moved. For a retired guide that has been replaced by a current guide on the same subject, that is a strong fit.

Keep a true hallucinated path as a 404 when no relevant equivalent exists. A made-up URL may look like a product page, a city page, a newsletter archive, or a knowledge-base article. Sending that visitor to the homepage does not answer the need that the assistant implied. It may also make it harder to see the recurring pattern in reporting.

Improve the not-found page for the visitor without pretending the missing page exists. Provide a clear explanation, a site search option if available, links to major sections, and contact or conversion routes that match the site. The page should help a human recover, while the status code and logging preserve the fact that the requested resource was missing.

Test every redirect after deployment. Confirm that the original URL returns the intended permanent status, the destination resolves successfully, and the destination is relevant to the original topic. Then annotate the audit entry with the date and reason. This allows the team to distinguish a repaired historical URL from an unresolved invented pattern at the next review.

The rule is not to eliminate every 404. The rule is to protect the visitor where a truthful destination exists, and preserve an honest signal where it does not. That is a better response to AI hallucinated links than building a broad catch-all redirect rule.

1. Use a permanent server-side redirect for a real page with a direct successor.
2. Do not redirect invented URLs to the homepage by default.
3. Improve the not-found experience without changing the missing-page evidence.
4. Verify status, destination relevance, and logging after every redirect.

How should brands measure the traffic impact over time?

Measure AI hallucinated links as a quality metric for AI-referred traffic. Track AI-referred sessions that land on 404 pages, the share of AI-referred landing URLs that fail, affected assistant, and outcomes after the not-found page. Keep this measure separate from ordinary organic-search 404s so the comparison remains meaningful.

The dated comparison from Ahrefs provides a practical benchmark for the problem, not a target to copy. On September 2, 2025, its clicked-URL analysis found a 0.43% average likely-404 rate for AI assistants and a 0.15% baseline for Google referrals. Its cited-URL analysis found ChatGPT at 2.38% versus 0.84% for Google top-20 URLs. Your own site may differ because its content history, redirects, instrumentation, and referral mix differ.

Use a before-and-after record for every remedial change. Before a redirect or not-found-page improvement, capture the requested URL, assistant source, sessions, status, and current destination. After the change, verify the status response and monitor whether the AI-referred path still receives traffic. This gives the team an evidence trail without claiming that a redirect altered an assistant's future citation behavior.

Tie the report to the broader citation program. Citation volume tells you how many times the brand is cited. Answer Presence tells you how broadly it appears across the question set. Citation Share shows the percentage of relevant answers that cite the brand. AI-referred 404 rate adds a different question: when a person follows a citation, does the destination work?

A clean report should state the limits. It should identify the date range, the referrers included, the rule used to detect 404s, and any missing server or conversion data. That makes the metric useful for decisions rather than a vague warning about AI traffic.

1. Report AI-referred 404 sessions and rate by assistant.
2. Compare the result with ordinary search referrals using the same measurement rule.
3. Log each redirect decision with evidence and verification.
4. Review the metric alongside Citation Share and Answer Presence.

What is the practical response plan for the next 30 days?

The practical response is to establish a narrow, repeatable audit rather than launch a broad redirect project. In the first week, create the AI-referral landing-page report and identify the highest-impact failed URLs. In the second week, classify those paths using site records. In the third week, redirect only the former real pages with relevant replacements and improve recovery routes on the not-found page. In the fourth week, verify the changes and publish a short measurement note.

This approach respects the uncertainty in the data. Ahrefs identifies a higher 404 rate among AI-assistant referrals, but its methodology also notes that some failures will be genuine 404s rather than hallucinations and that not every hallucinated link receives a click. Treat the study as a reason to inspect your own evidence, not as a license to make universal claims about any assistant.

The editorial implication is direct. A site should make important pages easy to identify, stable where possible, and well maintained when they move. That is not a way to game a model. It is standard care for people arriving from an answer, a search result, a social link, or a saved bookmark. Quality, coverage, and freshness remain the defensible path to earning and keeping citations.

For teams building an AI-search program, put the issue in the regular operating rhythm. Monitor cited destinations, measure traffic quality, and repair verified historical paths. This is Citation Engineering in practice: not chasing a clever trick, but making sure authoritative content remains usable when it is surfaced.

1. Week 1: build the AI-referral 404 report.
2. Week 2: classify the highest-impact paths from site evidence.
3. Week 3: deploy only relevant permanent redirects and recovery improvements.
4. Week 4: verify outcomes and record the dated result.

Sources

Source: Ahrefs

Ahrefs studied 16 million unique URLs and reported that AI assistants sent visitors to likely 404 pages at 0.43% on average, versus 0.15% for Google Search referrals. Ahrefs, 2025-09-02

Source: Google Search Central

Google explains that redirects tell visitors and Google Search that a page has a new location, and recommends permanent server-side redirects when a page has permanently moved. Google Search Central, 2026-04-14