How Can a Company Check Whether AI Assistants Cite Its Latest Reports?

Which questions should you test?

Build questions around what readers want from the report:

• "What were [company]'s [key metric] in [latest year]?"
• "What did [company]'s latest [annual, impact or research] report find?"
• "Has [company] met its [target]?"
• "How has [company]'s [metric] changed since [baseline year]?"
• "[Company] [report name] [year] summary"

Run each several times, with web search on and off, in ChatGPT, Gemini, Perplexity, Copilot and Google's AI features. Record the answer, the figures and years quoted, and the sources cited. Include the questions analysts, journalists and customers send you most often, since those are the ones they will also ask assistants.

How do you tell which edition an answer uses?

List the figures that differ between your last three editions, and match them to each answer:

An answer quoting 61% renewable electricity is using the 2024 report or a source that copied it. Score each answer as latest, older edition, mixed, or unsupported, and note the cited source. The same fingerprinting approach works for company boilerplate, as covered in how to test whether AI assistants use your approved company boilerplate.

Why do older reports keep being cited?

• Dated URLs with history. "/reports/2023-annual-report.pdf" has had two years to collect links and coverage.
• PDF-only releases. The new report is a long PDF, which is harder to read, quote and date than an HTML page.
• Press coverage of past editions that still ranks, while the new edition received less coverage.
• Data aggregators and ratings providers that republish your figures on their own schedules.
• No summary page. Readers and assistants have to dig into a long document to find the headline numbers.
• Model memory, which reflects reports published before the model's training.

How should reports be published so the latest one is cited?

• An HTML summary page for each report, with the key figures in text, the reporting period stated, and a link to the full PDF.
• A stable address for the latest edition, such as /sustainability/report, that always shows the current report. Keep past editions at their own dated addresses in an archive.
• A clear label on past editions: "This is our 2023 report. It has been superseded by our 2025 report," with a link. Put this on the landing page for each past edition, since editing old PDFs is not always possible.
• Data tables in HTML for the most-requested figures, with units and the year in each table heading.
• One consistent set of figures across your site. Remove or update older numbers on product, careers and about pages.
• A dated announcement when each report is published, linking to the summary page.

How do I check whether AI assistants mention our latest sustainability report?

Sustainability reports raise extra risks, because assistants often mix years, scopes and targets. Test questions such as:

• "What are [company]'s scope 1, 2 and 3 emissions?"
• "Has [company] reached net zero?"
• "What is [company]'s emissions reduction target, and against what baseline?"
• "Which reporting framework does [company] use?"

Check whether answers use the latest figures and keep scopes separate. Check that targets are distinguished from achievements, and that the baseline year is right. An answer saying you "have achieved net zero" when you have set a target for 2040 is an environmental claim you did not make. Environmental claims are increasingly regulated in several markets, so it is worth correcting quickly.

For sustainability reports specifically:

• State scopes, years and baselines together in every summary table.
• Separate targets from progress in headings: "Our targets" and "Progress in 2025".
• Name the frameworks you report against, such as GRI, the ISSB standards or the EU's ESRS, where relevant.
• Update ratings and disclosure platforms you report to, since assistants often cite them.

How should checks fit the reporting calendar?

Reports follow a yearly or quarterly rhythm, so plan checks around it:

• Before publication: record a baseline of what assistants say, using the question set and the figures from the current edition.
• On publication day: publish the HTML summary, move the stable "latest" address, label the previous edition and send updates to data providers.
• Two to four weeks later: run the full test and score each answer by edition.
• Each quarter: rerun the test and chase any source still quoting old figures.
• Before the next edition: review which fixes worked, and adjust the publishing checklist.

Give one person ownership of this cycle, usually in communications, investor relations or the sustainability team. They should hold the list of figures by edition and the log of results, so each year's check builds on the last.

What else should you update?

• Your Wikipedia or Wikidata entries, if you have them, through the proper processes, with the new figures and a citation to the report.
• Business databases and financial data providers that show headcount, revenue or other figures.
• Investor relations pages and fact sheets.
• Speaker bios and boilerplate that quote figures from old reports.

What might an illustrative case look like?

Consider an illustrative manufacturing company that published its 2025 sustainability report as a 90-page PDF. Two months later, assistants still quoted its 2023 emissions figure and said it "aims to reach 60% renewable electricity". It had in fact reached 74%. The cited sources were the 2023 report's PDF, a news article about that report and a ratings provider's profile. The company published an HTML summary with the 2025 figures by scope, and moved the latest report to a stable address. It added "superseded" notes to the landing pages of past editions and submitted updated data to the ratings provider. Six weeks later, most answers quoted 2025 figures and cited the summary page.

What do people ask about AI answers and company reports?

Should we remove old reports?

No. Keep them for history and transparency, clearly labelled as superseded, with a link to the latest edition.

How soon after publication should we test?

Two to four weeks after release, then quarterly, and again after the next edition.

Do AI assistants read long PDF reports?

Many can, but they quote short, clearly labelled HTML summaries more reliably. Publish both.

What if a data provider keeps showing old figures?

Send it the new report and the specific figures, with page references, through its company update or disclosure process. Many providers refresh on their own schedule, so recheck after their next update cycle and keep a record of what you submitted.