What's behind the link under an AI answer
AI search tools put sources under their answers, and the links look like receipts. Four independent audits, from 2023 to 2025, checked what is behind them. Every tool got some of it wrong, the gap between tools is wide, and almost nobody clicks through to check.
Does the source back the answer?
Three audits, three different tests. Compare tools within a panel, not across panels.
Do the citations back each statement?Four AI search engines, late February to late March 2023
- perplexity.ai 68.7%
- NeevaAI 67.6%
- Bing Chat 58.7%
- YouChat 11.1%
Share of generated statements fully supported by the citations given for them. Liu, Zhang and Liang 2023, Table 7.
Can it name the article an excerpt came from?Eight AI search tools, 200 excerpts each, February 2025
- Right
- Right but incomplete
- Partly wrong
- Wrong
- No answer
- Perplexity 37% wrong
- Copilot 40% wrong
- Perplexity Pro 45% wrong
- ChatGPT Search 67% wrong
- DeepSeek Search 67.5% wrong
- Gemini 76% wrong
- Grok-2 77% wrong
- Grok-3 94% wrong
Wrong counts partly and fully wrong answers, as the report does. Tow Center for Digital Journalism, 2025.
Do news answers have a significant sourcing problem?Four free assistants, 2,709 answers judged by journalists, 24 May to 10 June 2025
- Gemini 72%
- ChatGPT 24%
- Copilot 15%
- Perplexity 15%
A source that doesn't support the claim, no source at all, or a sourcing claim that is wrong or can't be checked. EBU and BBC 2025, p. 10.
How to read it
Each bar is one tool in one audit, split by what checkers found. The three audits asked different questions, so a tool's place matters only within its own panel: in 2023, whether each sentence was backed by its citations; in early 2025, whether a tool could identify the news article an excerpt came from; in mid-2025, whether answers about the news had serious sourcing problems.
- These are snapshots. The tools change their models every few months, and later versions may do better or worse.
- The 2025 breakdown by tool is the desk's count of the squares in the Tow Center's published chart. It reproduces every figure the report states: over 60% wrong overall, Perplexity 37%, Grok 3 94%, ChatGPT 134 of 200.
- The first two audits are US- and English-heavy. The third covered 18 countries and 14 languages.
Show the 2025 counts as a table
| Tool | Right | Right but incomplete | Partly wrong | Wrong | No answer | Wrong, % |
|---|---|---|---|---|---|---|
| Perplexity | 89 | 1 | 17 | 57 | 36 | 37 |
| Copilot | 16 | 0 | 14 | 66 | 104 | 40 |
| Perplexity Pro | 105 | 0 | 22 | 68 | 5 | 45 |
| ChatGPT Search | 56 | 10 | 21 | 113 | 0 | 67 |
| DeepSeek Search | 47 | 0 | 19 | 116 | 18 | 67.5 |
| Gemini | 1 | 2 | 34 | 118 | 45 | 76 |
| Grok-2 | 2 | 6 | 56 | 98 | 38 | 77 |
| Grok-3 | 10 | 0 | 71 | 117 | 2 | 94 |
Sources: Liu, Zhang and Liang, Evaluating Verifiability in Generative Search Engines (2023); Jaźwińska and Chandrasekar, AI Search Has a Citation Problem, Columbia Journalism Review (2025); European Broadcasting Union and BBC, News Integrity in AI Assistants (2025). Checked 7 November 2026.
Where the links point
Share of links going to each kind of site in Google's AI summaries and in its standard results. 68,879 searches by 900 US adults, March 2025.
- AI summary (first three source links)
- Standard results (first page)
- Wikipedia
- Government (.gov) sites
- News sites
- YouTube
8% of visits to a page with an AI summary ended in a click on a standard result, against 15% without one.
1% included a click on a link inside the summary.
26% ended the browsing session there, against 16% without a summary.
Source: Chapekis and Lieb, Google users are less likely to click on links when an AI summary appears in the results, Pew Research Center (2025). AI summaries appeared on 18% of these searches; their text was collected April 7 to 17, 2025. Checked 7 November 2026.
Log your own AI's sources for a month
The audits above are snapshots of other people's questions. This is how the desk checks the sources behind its own.
- Set up. List the tools you use, with the plan, the model or mode, and whether search is on. Write ten standing questions from your real work and ask them every Monday.
- Keep one row per citation: the date, the tool, the question, the exact sentence the citation sits under, the URL and the site.
- Open every link the same day. Note where it lands: the page itself, an error page, a homepage, a paywall, or a copy of the article on another site.
- Grade it against its sentence: backs it, partly, doesn't, or no page. Search the page for the key fact; a page on the right topic that lacks the fact counts as "doesn't."
- At the end of the month, count by tool: the share in each grade, the share of dead links, how much of the total the ten most-cited sites take, and how many of the sites you knew before.
Why it matters
AI answers now sit between a project and the people checking it, and the links under those answers are where a description gets confirmed or bent. The desk reads them as part of an audit: which pages the tools cite about a project, whether those pages say what the answer claims, and whether they exist at all. A link is a lead to check, not proof that anyone checked.
What language models say about a project