AI Answers Get Your Facts Wrong? Fix What Assistants Extract
Sometimes the problem is not that AI assistants ignore you but that they cannot use what you publish. Your brand is named, yet the answer gives an old price, a vague description of what you do, or cites a directory for facts that sit on your own site.
That usually means the facts are hard to lift out of your pages: hidden in images or scripts, spread across several paragraphs, or contradicted elsewhere on the site. This guide shows how to trace an answer back to your page, find why the fact was missed, and rewrite so the right detail is easy to take and quote.
What extraction problems look like in your results
Extraction problems show up as a pattern in tracked answers rather than as an error message. Look through the stored answers for prompts where you are named and check for these:
- Mentioned but never cited. The assistant knows your name but always links to a review site, a directory or a news article, even for facts about you such as your services or opening hours.
- Cited, but the detail is wrong. Your page is listed as a source, yet the answer states a price, a date or a feature that your page does not say, or no longer says.
- Vague descriptions. The answer calls you "a provider of business solutions" when your page makes a specific claim. The specific claim was probably not in a form the assistant could use.
- Half an answer. One part of a multi-part question is answered from your page and the rest from elsewhere, because the rest of your answer is on a different tab, page or PDF.
- A competitor wins on your own topic. Their page states the answer in a sentence; yours builds up to it over several paragraphs.
The AI Visibility Tracker helps here because it records whether you were cited, mentioned or invisible and surfaces wrong-attribution incidents. Those records are the starting point: each one points to a prompt, an answer and a page to examine.
If an assistant is inventing facts that appear nowhere, on your site or elsewhere, that is a different problem; see fixing AI hallucinations. If it credits your claim to someone else, or someone else's claim to you, see fixing wrong attribution.
What is known about how assistants use a page
The providers publish very little about exactly how their assistants choose and use passages, so be wary of anyone offering a precise formula. The general shape, as described in public documentation and visible in the answers themselves, is enough to act on.
- Retrieval. For questions that trigger a web search, the assistant runs one or more searches, using its own index, a search provider, or both, and fetches some of the pages it finds. Answers drawn only from a model's training data involve no fetch at the time of the question.
- Reading. The fetched page is reduced to text. What survives depends on the crawler: some run little or no JavaScript, and text inside images, video or embedded documents may not be read at all.
- Selecting. Only part of each page is used. Passages that directly address the question are more likely to be used than scene-setting, and a passage that makes sense on its own is easier to use than one that relies on the paragraphs around it.
- Writing. The assistant writes an answer in its own words, often combining several sources, and attaches citations. Paraphrasing is where a vague or ambiguous passage turns into a vague or wrong statement.
Every stage can lose your fact. The steps below check each stage in turn: can the page be fetched and read, is the fact present as text, is it clear on its own, and does the rest of your site agree with it.
Tracing an answer back to your page
Pick one prompt where the answer gets something about you wrong or vague, and work through it.
1. Find the fact on your site
Search your own site for the fact the answer should have used. Note every page where it appears and what each one says. If you cannot find it written anywhere in plain text, you have found the problem already.
2. Look at what is actually sent
A browser shows the page after scripts have run. A crawler that does not run scripts sees only the HTML the server sends. Fetch the page without a browser and search the response for the fact:
curl -sL -A "Mozilla/5.0 (compatible; extraction-check)" https://example.co.uk/pricing \
| sed -e 's/<[^>]*>/ /g' | tr -s ' ' | grep -i -o ".\{0,60\}per month.\{0,60\}"If the price appears in your browser but not in this output, it is being added by JavaScript and may be invisible to some assistants. In browser developer tools you can see the same thing by viewing the page source rather than the inspected elements, or by disabling JavaScript and reloading.
3. Check whether the crawler was allowed in
If the page is blocked for the relevant crawler in robots.txt, or refused by a firewall, nothing on it can be extracted, however well it is written. The crawler guides for ChatGPT and Perplexity cover how to confirm access from your logs.
4. Compare the wording
Put the answer's sentence next to the nearest sentence on your page. A close paraphrase of an out-of-date line means the page, or a copy of it, still says the old thing somewhere. A vague paraphrase of a long, hedged paragraph means the fact was never stated plainly.
The fix, step by step
Step 1
Line up the answer against your page
From your tracking results, pick the prompts where you are named but the facts are missing, vague or wrong. For each one, copy the relevant sentence of the answer and the list of sources, then find the page on your site that should have supplied that fact. Put the two side by side. Note whether your page was cited, whether it states the fact at all, and whether another page on your site says something different.
Step 2
Fetch the page the way a crawler receives it
Request the page with curl or view its source rather than the rendered
page, and search the raw HTML for the fact, such as the price, the service area or the opening hours.
If the fact only appears after JavaScript runs, after a click, inside an image or inside an embedded
widget, some assistants will never see it. Check robots.txt and your firewall too, because a blocked
page cannot be read at all.
Step 3
Bring hidden facts into the page text
Move every fact customers ask about into the HTML the server sends. Write prices, service areas, delivery times and key specifications as text rather than as images. Render tab and accordion content in the page and hide it with CSS instead of loading it on click. Add a written summary beside PDFs, videos and calculators. Keep the facts visible to human visitors as well, so the page reads the same for everyone.
Step 4
Make every key fact stand on its own
Rewrite each important fact as a sentence that makes sense out of context. Name the business or product instead of writing we or it, include units and currency, say what the figure covers and when it applies, and put the direct answer before any explanation. Brookvale Garden Rooms installs 3 by 4 metre garden offices from £14,500 including VAT and groundwork can be quoted safely; prices start from around fourteen and a half cannot.
Step 5
Remove contradictions across the site
Search the whole site, including old blog posts, landing pages and PDFs, for every version of the facts you just rewrote. Update or remove outdated prices, services and addresses, and redirect retired pages to their current equivalent. Then check the profiles you control elsewhere, such as your business listings and social profiles. When two of your own sources disagree, an assistant has to choose, and it may choose the older one.
Step 6
Use tables and lists for things people compare
Where customers compare options, such as packages, sizes, prices or delivery times, present the comparison as an HTML table with clear column headings, or as a list with one option per item. Keep the table in real HTML rather than an image. Add a sentence above the table saying what it compares and when it was last updated, so a single row quoted on its own still carries its meaning.
Step 7
Re-run the same prompts and compare wording
After the changes are live and the pages have had time to be fetched again, run the same prompts in the same assistants under the same conditions you used before. Compare the new answer sentences with your rewritten ones. Record whether the fact is now correct, whether your page is now the cited source, and whether other sources are still cited for facts about you. Keep the old answers so the before and after can be compared.
Rewriting passages so they survive being quoted
The same fact can be easy or hard to use depending on how it is written. These before and after pairs, for a fictional bookkeeping firm called Harlow Lane Bookkeeping, show the kind of change that helps.
| Before | Problem | After |
|---|---|---|
| We have been helping local businesses for many years with all their needs. | No name, no service, no place, no fact. | Harlow Lane Bookkeeping has provided bookkeeping and VAT returns for sole traders in Harlow since 2014. |
| Our packages are really affordable and flexible, just get in touch. | No price at all. | Monthly bookkeeping for sole traders costs £85 a month plus VAT, including quarterly VAT returns. |
| It integrates with most popular tools. | "It" and "most" cannot be quoted. | Harlow Lane works with clients using Xero, QuickBooks and FreeAgent. |
| Turnaround is quick. | No measure. | Harlow Lane files VAT returns at least five working days before the deadline. |
Patterns worth copying
- Answer first. Put the direct answer in the first sentence under a heading, then add the detail and caveats.
- One fact per sentence where precision matters. Long sentences with several clauses are where paraphrasing goes wrong.
- Say when. Prices, availability and rules change. "Prices from April 2026" tells a reader, and an assistant, how current a figure is.
- Headings that name the topic. "Prices for sole traders" helps a passage to be matched to a question; "Our approach" does not.
None of this means writing for machines at the expense of people. Clear, specific, dated statements are what customers want too.
Platform notes
WordPress
Page builders often put tab, accordion and slider content in the HTML already, which is fine. Check plugins that load content on click, such as some pricing table and FAQ plugins, by viewing the source. Prices saved as images inside the media library are a common cause of missing facts.
Shopify
Product descriptions are in the HTML by default, but apps that add size guides, delivery estimates or extra tabs may load them with JavaScript. View the page source for a product and search for the text the app adds; if it is missing, look for a theme setting or metafield that renders the same information in the template.
JavaScript frameworks
Single-page apps built with React, Vue or similar frameworks may send almost no text in the initial HTML. Server-side rendering or static generation, available in frameworks such as Next.js and Nuxt, puts the content in the HTML. See fixing JavaScript-only content for the details.
Static sites
Static HTML is usually easy to read. The typical problems are facts inside images, stale copies of old pages left on the server, and price lists kept only as PDFs.
Keeping facts extractable
- Keep one source of truth for key facts. Prices, opening hours, service areas and contact details should come from one place in your CMS and flow into every page, rather than being typed separately on each.
- Check the raw HTML after redesigns. A new theme or plugin can move content behind scripts without anyone noticing in the browser.
- Date your key pages honestly. Show when prices and terms were last reviewed, and update the date only when the content changes.
- Retire old content properly. Redirect or update outdated pages rather than leaving them live and unlinked.
- Watch the answers. Keep tracking the prompts that exposed the problem, so you notice if a fact starts going wrong again.
1. Why does ChatGPT quote an old price from my website?
Usually because an old price is still published somewhere it can be found, such as an outdated blog post, a PDF price list, a cached copy on another site or a directory listing. It can also happen if the answer was drawn from what the model learned in training rather than a live search. Search your own site and your business profiles for every version of the price, update or remove the old ones, and state the current price with a date on one clear page.
2. Can AI assistants read content inside tabs and accordions?
If the content is in the HTML the server sends and is simply hidden with CSS until clicked, crawlers that read the HTML can see it. If the content is fetched by JavaScript only when someone clicks, it is not in the initial HTML and crawlers that do not run scripts will not see it. View the page source or fetch the page with curl and search for the tab text to find out which kind you have.
3. Do AI assistants read text in images?
You should not rely on it. Some systems can analyse images, but the crawlers that gather pages for search-based answers mainly work with text, and alt text is usually too short to carry a full price list or specification. If an image contains facts customers ask about, such as prices, opening hours or sizes, repeat those facts as ordinary HTML text on the same page. That also helps visitors using screen readers or slow connections.
4. Why is a directory cited instead of my own site for facts about my business?
The directory may state the fact more plainly, in a form that is easier to lift, or your own page may hide it in an image, a script or a long paragraph. It is also possible that the assistant could not fetch your page at all because of a robots.txt rule or firewall. Check access first, then make sure your own page states the same fact clearly as text, and correct the directory listing if it is wrong.
5. Does adding schema markup fix extraction problems?
Structured data can help systems identify facts such as your organisation name, address or product price, but it is not a substitute for clear visible text. The providers have not said that their assistants rely on schema markup for answers, so treat it as a supporting signal. Keep the markup accurate and matching what the page says, and make sure every important fact is also written plainly in the page text where people and crawlers can both read it.
6. How long should an answer paragraph be for AI assistants?
There is no published ideal length. What matters is that the first sentence under a heading answers the question directly and makes sense on its own, naming the business or product and including units, prices and dates where relevant. Supporting detail can follow in as many sentences as the reader needs. A short, specific opening followed by explanation serves both a person skimming the page and an assistant looking for a quotable passage.
7. Will rewriting my pages change what assistants say straight away?
Not usually. Assistants that search the web need to fetch the updated page again, and answers drawn from a model's training data may not change until a later model is released. Some assistants may also keep using other sources that repeat the old information. Make the changes, correct any outside sources you control, then re-run the same prompts periodically and compare the answers with those you stored before, rather than checking once the next day.
The AI Visibility Tracker queries ChatGPT, Claude, Perplexity, Gemini, Copilot and Brave Leo with your target prompts, records whether your brand is cited, mentioned, recommended or invisible, and surfaces wrong-attribution incidents so you know which answers to trace.
Open AI Visibility Fixes →