Ask an AI engine a question about your brand and you get one confident paragraph. It might state your price, your founding year, the awards you won or the tools you connect with. What it rarely tells you is how many independent pages that paragraph actually rests on.
Often it is only one or two. When a fact about your brand rests on a single page, that page decides what people are told. If the page is wrong, the AI repeats the mistake. If the page is deleted, the answer changes. This guide shows how to find those weak spots with a simple "source concentration audit," and how to fix them.
What a "single-source" AI answer is
A single-source answer is an AI answer where a key fact about your brand is backed by only one website (one domain). The domain might be your own site. It might be a review blog, a directory, or an old press article.
Two words will come up a lot in this guide:
- Citation: a link the AI shows next to or below its answer to say "this is where the information came from."
- Corroboration: when several separate sources say the same thing. If three unrelated websites all list the same price, that price is corroborated.
Single-source answers are not always wrong. Your own pricing page is the best source for your price. The problem is that an answer leaning on one page has no backup if that page is wrong or changes.
How AI engines choose the pages they cite
Here is what the main engines say about how they pick and show sources.
| Engine | What the platform says | What it means for this audit |
|---|---|---|
| Google AI Overviews and AI Mode | Both "may use a 'query fan-out' technique," which Google describes as "issuing multiple related searches across subtopics and data sources" to build a response. Google says this lets it find more supporting pages and show "a wider and more diverse set of helpful links." A page must be indexed and eligible to show with a snippet to be used.1 | One question can trigger several hidden searches, each pulling in different pages. The links you see are a sample of what was read. |
| ChatGPT search | "Responses that use web search may include citations," and a Sources button, when available, lists cited sources and other relevant links. OpenAI also warns that "Search results and citations can be incomplete, outdated, or incorrect."2 | If the only cited page is outdated, the answer can be too. |
| Perplexity | "Each answer includes numbered citations linking to the original sources."3 | Numbered citations show which claim is tied to which page, so it is a good place to start. |
None of these documents promises that a cited page fully supports every sentence near it, or that the most visible link shaped the answer most. That gap is where single-source risk hides.
Three ways a single-source answer can hurt you
1. One wrong or stale page gets repeated as fact
Suppose a review blog lists your old price or a made-up claim. If no other page covers that fact, the AI has nothing to compare it with. So it may repeat the error confidently, with a citation that makes it look checked.
Independent research shows that AI answers already struggle with sources even when plenty exist:
| Study | What it measured | Key finding |
|---|---|---|
| Stanford researchers Liu, Zhang and Liang (2023) | Whether citations in four generative search engines (Bing Chat, NeevaAI, perplexity.ai and YouChat) backed up the sentences they were attached to | On average, only 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations supported their sentence.4 |
| Tow Center for Digital Journalism, Columbia Journalism Review (March 6, 2025) | 1,600 queries across eight AI search tools, asking each to identify the source of a real news excerpt | The tools gave incorrect answers to more than 60% of queries. Perplexity was wrong on 37%. Grok 3 was wrong on 94%. Several tools cited syndicated copies of articles instead of the original publisher.5 |
| EBU and BBC News Integrity in AI Assistants study (October 22, 2025) | More than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity, judged by 22 public service media organisations in 18 countries and 14 languages | 45% of answers had at least one significant issue. 31% had serious sourcing problems, meaning missing, misleading or incorrect attributions. 20% had major accuracy issues, including outdated information.6 |
| The New York Times with AI startup Oumi, reported by Search Engine Land (April 7, 2026) | 4,326 Google searches using the SimpleQA fact-checking benchmark | AI Overviews were accurate 91% of the time with Gemini 3, up from 85% with Gemini 2. But 56% of the correct answers in February were "ungrounded," meaning the linked sources did not fully support the answer. Google disputed the method, saying the benchmark does not reflect what people actually search.7 |
None of these studies looked at brand facts directly. They tested news and general knowledge questions. But they point the same way: the link between an AI sentence and its cited page is often loose. If your brand fact sits on one page, there is no second source to catch that.
2. One page disappears and your answer changes
Pages vanish all the time. A directory deletes old listings, or your own team moves a page without a redirect.
If that page was the only source for a fact, the AI has to find something else. It might fall back to an older, worse page, or drop the fact. You see the answer change even though nothing in your own marketing changed.
3. Citation dashboards may credit the wrong page
Most AI visibility tracking counts which domains appear in the source links (sometimes called "source chips") next to an answer. That measures what was shown, not what did the work.
Query fan-out means one answer can be built from several hidden searches.1 And cited pages often do not fully support the sentences near them.47 So your dashboard may say "your site was cited" while the wording came from a third-party page that was never shown. Or a review site gets credit for a fact it never stated.
The audit below checks both.
The source concentration audit, step by step
The first run takes about half a day. A monthly rerun takes an hour or two. You need a spreadsheet and the AI engines your customers use.
- List your key facts. Write down 6 to 10 facts that matter for buying decisions. Typical ones: price or price range, free plan or trial, founding year, headquarters, main integrations, awards, certifications, and the one-line description of what you do.
- Write 20 prompts. About 8 that name your brand ("How much does [brand] cost?"), 8 category questions ("best invoicing tool for freelancers") and 4 comparisons ("[brand] vs [rival] pricing").
- Run each prompt on each engine. Use Google AI Mode or AI Overviews, ChatGPT search and Perplexity. That gives you 60 answers. Use a logged-out or clean session where you can, so past chats do not shape the result.
- Record every cited URL. For each answer, paste the full URL of every citation into the sheet, along with the domain.
- Tag the key fact in each answer. Note which of your key facts the answer states, and whether it is correct, stale (once true, now out of date) or wrong.
- Check support, page by page. Open each cited page and look for the fact. Mark the page "supports" only if the fact is actually on the page. This is the step most teams skip, and it is the one that catches the dashboard-credit problem.
- Count independent supporting domains. For each answer, count how many separate domains support the key fact. Two pages on the same site count as one. A syndicated copy of the same article on another site also counts as one, because it is the same source wearing a different address. The Tow Center study found AI tools citing syndicated copies instead of originals, so this matters.5
- Find the "shaping" page. When an answer uses wording you recognise, search for that exact phrase. If it comes from a page that is not cited, note it in an "uncited source" column.
- Calculate the scores (see the next section).
- Save the raw sheet. AI answers change between runs, so one run is a snapshot. Keep each month's sheet to compare.
Four scores that show how concentrated your answers are
Only count answers that state at least one of your key facts. Call that number your "fact answers."
- Single-source rate: fact answers where the key fact is supported by only 1 domain, divided by all fact answers.
- Top-domain share: citations going to the single most-cited domain, divided by all citations in fact answers.
- Owned-only rate: fact answers where the only supporting domain is your own site, divided by all fact answers. A high number here means third parties are not backing you up.
- Error rate: fact answers with a stale or wrong key fact, divided by all fact answers.
For example, if 54 answers state a key fact and 31 of them rest on one domain, the single-source rate is:
31 ÷ 54 = 57%
These are Ripplix's working bands for reading the scores. They are rules of thumb from practice, not an industry standard, so adjust them to your market.
| Score | Low risk | Watch | High risk |
|---|---|---|---|
| Single-source rate | Under 25% | 25% to 50% | Over 50% |
| Top-domain share | Under 30% | 30% to 45% | Over 45% |
| Owned-only rate | Under 15% | 15% to 30% | Over 30% |
| Error rate | Under 5% | 5% to 10% | Over 10% |
Worked example: a composite brand, before and after
Illustrative composite: "Quillsby" is an invented invoicing app for freelancers. The brand, the domains and every number below are made up to show how the audit works. They are built from patterns seen across many brands, not taken from a real client or a real study.
Quillsby ran the 20-prompt audit across three engines, giving 60 answers. 54 of them stated at least one key fact.
The "before" picture
| Key fact | Fact answers | 1 domain | 2 domains | 3+ domains | Wrong or stale | Main source |
|---|---|---|---|---|---|---|
| Price | 14 | 9 | 3 | 2 | 7 | An old review post on ledgerlane-reviews.example listing a $12 plan Quillsby dropped |
| Free plan | 9 | 6 | 2 | 1 | 2 | Same old review post |
| Integrations | 10 | 4 | 4 | 2 | 1 | Quillsby help center |
| Award ("Best Freelancer Tool 2025") | 6 | 6 | 0 | 0 | 0 | Quillsby's own press page only |
| Founding year and HQ | 8 | 4 | 2 | 2 | 1 | A directory listing with the wrong city |
| What it does (one-liner) | 7 | 2 | 3 | 2 | 0 | Mixed |
| Total | 54 | 31 | 14 | 9 | 11 |
Across those 54 answers there were 142 citations. 61 pointed to quillsby.example, its own site. The old review post's domain had 24.
- Single-source rate: 31 ÷ 54 = 57% (high risk)
- Top-domain share: 61 ÷ 142 = 43% (watch)
- Owned-only rate: 17 ÷ 54 = 31% (high risk; 17 of the 31 single-source answers rested only on Quillsby's own site)
- Error rate: 11 ÷ 54 = 20% (high risk; 9 of the 11 errors traced back to the one old review post)
The support check (step 6) also turned up 7 answers where Quillsby's own site was cited, but the cited page did not contain the fact. In 5 of those, the wording matched the old review post, which was not shown as a source. A dashboard would have counted all 7 as wins for Quillsby. In fact, the review post shaped them.
The award row is worth a closer look. Every answer that mentioned the award cited only Quillsby's press page. The award was real, but nothing independent confirmed it. And if Quillsby moves that press page, the award vanishes from AI answers.
The fixes, in priority order
| Problem found | Fix | Why it helps |
|---|---|---|
| Old review post lists a dropped $12 plan and wrong free-plan terms | Emailed the site owner with the current pricing page link and a short changelog. The post was updated within two weeks. | Fixes the root page for every engine. |
| Price and free plan only clear on one owned page | Rewrote the pricing page so each plan, price, currency and "last updated" date sits in plain text, not only inside images or toggles. Added the same facts to the help center FAQ. | Easier to read and quote, with a backup page. |
| Award backed only by the brand's own press page | Asked the award organiser to list Quillsby on its public winners page. Added the award to the brand's profiles on two major software review sites. | Moves the claim from "the brand says so" to "the organiser says so." |
| Directory shows wrong HQ city | Claimed and corrected the listing. Checked the five other directories that appeared in citations. | Removes a wrong source. |
| Integrations only on help center | Got listed in the app marketplaces of the three main integration partners. | Adds independent domains confirming the fact. |
| Old pricing URL had no redirect | Added a permanent redirect from the old URL to the current pricing page. | Stops a weaker page filling the gap. |
The "after" picture, six weeks later
Quillsby reran the same 20 prompts on the same three engines. This time 56 answers stated a key fact, with 171 citations.
| Score | Before | After | Band now |
|---|---|---|---|
| Single-source rate | 57% (31 of 54) | 27% (15 of 56) | Watch |
| Top-domain share | 43% (61 of 142) | 30% (52 of 171) | Watch, at the edge of low |
| Owned-only rate | 31% (17 of 54) | 13% (7 of 56) | Low |
| Error rate | 20% (11 of 54) | 5% (3 of 56) | Watch, at the edge of low |
| Answers with 3+ supporting domains | 9 | 22 |
Two caveats. AI answers vary between runs, so part of any change can be noise. And the team made several fixes at once, so it cannot say which fix caused which change. The audit shows the answers became less fragile. It does not prove cause and effect.
A scoring sheet you can copy
Set up one row per answer in your spreadsheet with these columns:
| Column | What to enter |
|---|---|
| Prompt ID | P01 to P20 |
| Engine | Google AI Mode, ChatGPT, Perplexity |
| Run date | The date you ran the prompt |
| Key fact stated | Which fact from your list, or "none" |
| Fact status | Correct, stale or wrong |
| Cited URLs | Every citation, full URL |
| Supporting domains | Domains whose cited page actually contains the fact (count syndicated copies once) |
| Support count | Number of independent supporting domains |
| Owned only? | Yes if the only support is your site |
| Cited but not supporting | Cited domains that do not contain the fact |
| Uncited source | A page that matches the answer's wording but was not cited |
| Fix owner | Who will act on it |
Then add one summary block per run:
| Summary score | Formula | Your result |
|---|---|---|
| Fact answers | Count of rows where key fact is not "none" | |
| Single-source rate | Rows with support count = 1 ÷ fact answers | |
| Top-domain share | Citations to most-cited domain ÷ all citations in fact answers | |
| Owned-only rate | Rows with "owned only" = yes ÷ fact answers | |
| Error rate | Rows with status stale or wrong ÷ fact answers | |
| Hidden-shaping rate | Rows with an uncited source ÷ fact answers |
What to do about single-source answers
The goal: every important fact is stated correctly on your own site and confirmed by at least two or three independent sites.
- Fix the original wrong page first. If one bad page is behind most of your errors, correcting it beats publishing ten new pages around it. Contact the site owner with the correct fact and a link to proof.
- State facts plainly on your own pages. Put prices, plan names, dates, locations and awards in normal text. Avoid hiding them in images, tabs that need clicking, or PDFs. Add a visible "last updated" date. Google says a page has to be indexed and eligible to show with a snippet before it can support an AI feature, so basic crawlability still counts.1
- Keep one main page per fact. If your price appears on five owned pages, they will drift apart. Pick one main page and link other pages to it.
- Claim and update third-party profiles. Review sites, software directories, app marketplaces, business listings and industry association pages are all independent domains. Each correct profile is one more source that agrees with you.
- Ask the original source to publish. For awards, certifications and partnerships, the most trusted confirmation comes from the body that gave it. A winners page on the award organiser's site is worth more than your own press release.
- Redirect old URLs. When you move or delete a page that AI engines cite, add a permanent redirect to the new page so the fact does not fall into a gap.
- Do not fake corroboration. Making up third-party sites, reviews or "independent" articles to back your own claims is manipulation. On May 15, 2026, Google clarified that its spam policies apply to generative AI responses in Google Search.8 Its spam policy now names "attempting to manipulate generative AI responses in Google Search" as part of what spam means.9
See which of your AI answers rest on a single page
Ripplix tracks the prompts your customers ask across AI engines and records every source each answer cites, so you can spot facts that depend on one domain before a stale page spreads. Start with a free report on your brand.
Get your free AI Visibility Report →
Sources:
- Google Search Central: AI features and your website (last updated December 10, 2025)
- OpenAI Help Center: ChatGPT search (accessed October 19, 2026)
- Perplexity Help Center: How does Perplexity work? (accessed October 19, 2026)
- arXiv: Liu, Zhang and Liang, Evaluating Verifiability in Generative Search Engines (April 2023, revised October 2023)
- Columbia Journalism Review, Tow Center: AI Search Has a Citation Problem (March 6, 2025)
- European Broadcasting Union: AI's systemic distortion of news is consistent across languages and territories (October 22, 2025)
- Search Engine Land: Report on the New York Times and Oumi analysis of AI Overviews accuracy (April 7, 2026)
- Google Search Central: Clarifying that spam policies apply to generative AI responses in Google Search (May 15, 2026)
- Google Search Central: Spam policies for Google web search (last updated August 28, 2026)
