Which AI models cite non-resolving URLs?

We regularly publish aggregated, anonymised figures from our platform. Here is a simple question our data can examine: when an AI cites a source, does that link resolve when checked? And which models show the most non-resolving cases in this dataset?

In short

The study reports 26,435 cited URLs from 53 models and 70 cases that did not resolve when checked. Among the table’s seven models, Perplexity Sonar Pro has the highest published rate, 1.05%, versus 0.04% for Perplexity Sonar. The counting rule is unspecified: these published rates do not measure risk per citation.

The scope

Between September 2025 and July 2026, the platform analysed 7,774 AI answers from 132 audits, produced by 53 distinct models. They contained 26,435 URLs cited as sources. At the time of checking, 70 URLs did not resolve, across 23 brands. The source publishes an overall rate of 0.26%, subject to the counting caveat below.

The published criterion counts a URL cited as a source that returns HTTP 404 or shows a soft 404 when checked. Mere inability to access a URL is therefore not enough. This does not prove that a page was never published or that the model invented it. The source does not specify whether the counts represent distinct URLs or occurrences, or whether the same rule applies to the numerator and denominator. We retain the rates as published figures, not as estimates of risk per citation.

The ranking: URL non-resolution rate at the time of checking

The table presents only the seven models shown and groups the 70 reported cases. Its “Cited URLs” column totals 23,654, versus 26,435 for the whole study.

ModelNon-resolving URL rateNon-resolving URLsBrands affectedURLs cited
Perplexity Sonar Pro1.05%2192,002
Anthropic Claude Sonnet 4.60.62%951,449
OpenAI GPT-5 Mini0.44%24135,436
Anthropic Claude Haiku 4.50.19%834,198
OpenAI GPT-50.16%321,848
Google Gemini 2.5 Flash0.12%211,656
Perplexity Sonar0.04%337,065

How to read it: the Perplexity Sonar Pro row reports 21 non-resolving URLs, 2,002 cited URLs, 9 brands and a 1.05% rate. Results with 3 cases or fewer remain indicative.

Finding 1: The "search engine" paradox

Across the seven rows in the table, Perplexity Sonar Pro has the highest published rate (1.05%) and Perplexity Sonar the lowest (0.04%). The gap is about 25×. This comparison describes this dataset and these specific models; it does not demonstrate the cause of the gap or a durable vendor ranking.

Finding 2: The highest raw count in the table

In this table, GPT-5 Mini has the largest raw count of non-resolving URLs (24, across 13 brands), with a 0.44% rate. The available data does not provide query volumes by model, so it cannot explain that total by frequency of use or any other cause.

Finding 3: Every tabled model has a detected case

Each of the seven models in the table recorded at least one URL that did not resolve when checked. That finding is limited to those seven models and does not support a conclusion about the other models among the 53 studied. In a separate barometer covering 699 addresses, 8 addresses received human visits. Whether they are among this study’s 70 URLs has not been established.

What it changes for a brand

Two concrete takeaways. First, the model you are visible in matters as much as being visible: a citation that leads to a non-resolving URL can associate you with an unavailable page. Second, measuring per model and checking cited URLs documents this phenomenon from the sources actually displayed.

The limits of these figures

The data comes from brands that chose to audit themselves and is not representative of all companies. Results with few cases remain indicative. The table covers seven models; its published cited-URL total is 23,654, versus 26,435 for the study.

Why the rate matters more than a raw total

A rate cannot be read without its denominator. The “Cited URLs” column reports 2,002 for Perplexity Sonar Pro, 5,436 for GPT-5 Mini and 7,065 for Perplexity Sonar. We reproduce those values without relabelling them as citation occurrences: the source does not confirm that unit.

GPT-5 Mini has the table’s highest published non-resolving URL count: 24. That count describes cases reported by the source, not clicks or user exposure.

Perplexity Sonar has both the table’s highest published cited-URL total, 7,065, and its lowest published rate, 0.04%. That alone cannot rank model reliability: the counting rule and its consistent application to both terms of the rate must first be established.

What the results establish

The table reports non-resolving URLs observed in the study. It attributes neither an origin nor a cause to any individual URL.

How to read the table without overclaiming

The seven rows report published counts and rates. Rows with 3 cases and cited-URL totals of 1,848 and 7,065 do not by themselves establish the relative statistical precision of the results: the counting unit is undocumented. The table groups 70 cases and a published total of 23,654 cited URLs, not the study’s full 26,435.

The table is not a score for the overall quality of an answer. A model can produce a useful answer while citing a URL that returns a 404 or shows a soft 404 when checked. Conversely, a URL that resolves does not by itself prove that its content properly supports the claim beside it. This study isolates the result of a check at a particular time, not proof that a page never existed.

For wider context, read our 2026 AI visibility barometer, which is separate from this study and covers 699 addresses. Our guide to non-resolving URLs sets out the precautions to take when an address attributed to a brand cannot be found at the time of checking.

What to do when a phantom URL appears

Start by retaining evidence: the displayed answer, the exact URL, the model, the date, and the result of the check. Without those details, a team cannot distinguish a typing mistake, an old page that was removed, or an address whose origin remains unknown. Then check whether the URL resembles a historical page, a spelling variant, or a destination that the site should redirect. Do not assign the non-resolution to a cause before checking the site and the link.

Next, look at frequency. One missing link is a signal to document. Repeated occurrences across answers or models justify a fuller examination of brand presence and of the pages AI systems cite. Model-by-model verification is more useful than one screenshot because it reveals whether the issue repeats and in what response context.

Finally, do not turn the finding into an opposite promise. Publishing more pages does not guarantee that a model will stop citing non-resolving URLs. The practical response is to make important pages accessible, coherent, and verifiable, then measure the sources that are actually displayed. Useful AI visibility is built on evidence and checks, not a slogan.

The AGS methodology describes the broader measurement framework, and the glossary defines AI hallucination.

Questions and answers

Which AI model invents the most URLs?

In our data (26,435 verified cited URLs), Perplexity Sonar Pro shows the highest rate: 1.05% of its citations point to a page that doesn't exist, across 9 brands. Conversely, Perplexity Sonar — from the same vendor — is the most reliable of the panel (0.04%). Reliability depends on the model's configuration, not the brand.

Do all AIs invent URLs?

Yes. Every model family in our study (OpenAI, Anthropic, Google, Perplexity) produced at least one invented URL. No model reaches 0%. The overall average rate is 0.26%, roughly 1 cited URL in 380.

How do you detect a URL invented by an AI?

A URL is counted as 'invented' only if the model cited it as a source and it returns an error (404 or soft-404) at HTTP verification time. The criterion is binary and verifiable: the page exists, or it does not.

Why is an invented URL dangerous for a brand?

Because users trust the source the AI displays. A phantom address attributed to your brand can receive real human visits that land on an error page — a reputation blind spot. In our 2026 barometer, 8 of these invented addresses had received real traffic.

About the author

Davy Abderrahman

Founder of

As a specialist in visibility in AI answers (AEO/GEO/LLMO), I help agencies and consultants measure and analyze their clients’ presence on ChatGPT, Claude, Gemini, Perplexity and other AI answer engines.

AEO GEO LLMO AI visibility AI audits
How do your brands appear in AI answers?

Agencies, consultants and in-house teams: compare the answers generated for your own scenarios, languages and selected models. Our infrastructure, your scenarios, your expertise.

Was this article helpful?

- (0 votes)