The AI Page Refresh Checklist: Make a Page Citable Without Inventing Facts

Where the 40% Came From

The figure comes from the GEO paper, published at KDD 2024 and the paper that coined the term. Its setup was a generative engine built on GPT-3.5 that answered each query from five retrieved sources. The researchers rewrote one of those sources with each method and measured how much of the answer it then earned, weighted by where in the answer it appeared. They report that their top three methods, Cite Sources, Quotation Addition and Statistics Addition, improved that measure by 30 to 40%.

The prompts are public

The paper's appendix says all of its prompts are in its public repository and that GPT-3.5 ran every experiment, including the rewriting. In that repository, the statistics method asks the model to add statistics "(even if hypothetical)". The quotation method asks for quotes "even though fake and artificial", in a prompt that also asks them not to mislead. The citation method says "You may invent these sources but ensure they sound plausible and do not mislead the reader". The repository's method list shows those are the prompts its quotation, citation and statistics runs call.

What that means for the number

The study supports a narrower claim than the one people quote: adding statistics and quotes, true or not, raised a source's share of one simulated engine's answers. Whether true figures do the same on a live page is a question it never asked, and so is what an invented statistic costs a business once a customer or a journalist checks it.

What the Study Does Show

Read the paper's own table rather than its headline and the useful result appears. Fluency Optimization asked the model to make the text read better without altering the core content. It scored 24.7, within 0.1 of Cite Sources at 24.6 and close to Statistics Addition at 25.2, against 19.3 for the untouched source.

The ordering among those close methods is not stable. Over five random seeds (the paper's Table 6), fluency scored 24.4 against 25.3 for Cite Sources and 25.5 for Statistics Addition, with deviations of ±0.6 to ±1.2. On a 200-query Perplexity test, Statistics Addition scored 26.2 and Fluency Optimization 26.0, against 24.1 untouched. Quotation Addition led both tests (27.2 and 29.1), and its prompt asked for quotes "even though fake and artificial".

The lesson to take

A model reading several sources favours text it can lift cleanly: specific, well ordered, quotable. You can produce that from what the page already knows, by restructuring and clarifying, and you can strengthen it with real figures and real quotes where you have them. What you cannot do is borrow the study's method, because on a live page an invented number is a claim your business has made.

The 12-Point Refresh Checklist

Run it on one page at a time, starting with a page AI already cites or one that answers a question you track and lose. The groups go in order: fix the structure, then the facts, then the markup that describes them.

AI PAGE REFRESH CHECKLIST

STRUCTURE: say what the page already says, better
  1. The answer comes first. The page's conclusion is in its first two or
     three sentences, in words a model can lift without the rest.
  2. Headings are the questions buyers ask, ideally the ones you track.
  3. One idea per paragraph, short enough to quote on its own.

FACTS: only what you can stand behind
  4. Every number carries its source and its date, next to it.
  5. A missing fact becomes a visible note for the person who owns it,
     never a plausible number.
  6. Quotes come from real people who said them, with name and role.
  7. No time-boxed claim you will not maintain ("this quarter",
     "currently"), unless it has a review date.
  8. Comparisons name the alternatives a buyer would weigh, fairly.

MARKUP: mirrors the page, never adds to it
  9. FAQ answers are taken from the page itself.
 10. Schema describes what is visible on the page and nothing more.
 11. dateModified changes only when the content substantively changed.

AFTER PUBLISHING
 12. Re-audit the page, then re-measure the questions it targets across
     at least 2 tracking runs before calling the change a result.

Structure (1 to 3)

These are the Fluency Optimization lesson applied by a person who knows the subject. Put the conclusion where a model looks first, phrase headings as the questions a buyer types, and keep paragraphs short enough to stand alone. The questions to use are the ones you already track; in Ranqo they live on the Prompts page, and the optimizer reads up to five of the ones you do not win yet. If the page has no brief behind it, start from our content brief template.

Facts (4 to 8)

This group is where refreshes go wrong, because the easiest way to make a page look more citable is to make it more specific than it is. Item 5 is the one to hold to: when the page would be stronger with a figure you do not have, the refresh records the gap and names who can fill it. Quotes follow the same rule as numbers, and comparisons earn trust by naming the alternatives a reader would weigh anyway.

Markup (9 to 11)

An FAQ and structured data should restate the page, not extend it. Google says plainly that no special schema is needed to appear in AI Overviews or AI Mode, and its guidance on AI-written content extends the fact-check to metadata and structured data. Treat schema as verification, not amplification. For dates, Google's advice is older than AI search: don't artificially freshen a page without adding significant information.

After publishing (12)

A refresh is a hypothesis until the answers move. AI answers shift from run to run, so one reading after publishing says little. Our noise measurements found an overall score moving 14.5 points or more between consecutive runs about one time in twenty. Re-audit the page so you know the change landed, then judge it by the specific questions it was meant to win, over at least 2 runs.

Where Refreshes Go Wrong

Asking a tool to add sources

A model asked to support a claim can supply support that does not exist. In a 2023 test of model-written literature reviews, 55% of GPT-3.5's citations and 18% of GPT-4's were fabricated. The newer model did better, but neither was clean, and we see the same pattern in page rewrites: the more a prompt rewards specificity, the more likely a missing fact gets filled in. Google's position is that it is critical to manually factcheck AI-generated content before publishing.

Faking freshness

Models do prefer recent dates. In a SIGIR-AP 2025 study, injecting fresher dates into passages moved the average publication year of an LLM reranker's top ten forward by up to 4.78 years and moved single passages by as many as 95 ranks. That is exactly the incentive to bump a date without changing the page, and exactly what item 11 rules out: a date is a claim about the content, and a reader who finds last year's figures under this month's date stops trusting the rest.

Claims with an expiry date

"Currently", "this quarter" and "as of this year" read well on the day you publish and become false on a schedule nobody tracks. A refreshed page tends to sit untouched for a long time afterwards, so either remove the time box or give the claim an owner and a review date.

Refreshing at scale

Running every page on a site through the same rewrite produces many pages that changed without improving, which is close to how Google's spam policies describe scaled content abuse. A short list of pages that matter, each checked by someone who knows the subject, beats a site-wide pass.

How Ranqo Does It

The Content Optimizer follows the same rule for the text it rewrites. The failure it guards against is concrete: ask a model to make a page about import duties more citable and it will happily supply a court ruling and a duty rate the page never mentioned. The optimizer is told to invent nothing, and when a page is missing a fact that would make it more citable, to say so in the description of the change, as something for you to add and verify, instead of writing it into the new text. That is an instruction to the model, not a check on its output, which is why the facts and markup groups of the checklist stay yours.

Pick the page

With the address field empty, Start from a page suggests two lists: AI cites these, your pages AI already cites, most cited first, and Audited, pages you have run through Page Audit, lowest score first. The optimizer only rewrites pages on your own site.

Read the changes

Each pass returns its changes labelled Rewrite, Add or Restructure, each with its After text and, where it replaces existing wording, the Before text above it. It also writes an FAQ for the page and schema markup for its head. It reads the page's audit scores when it has them and up to five of the tracked questions you do not win yet (your weakest overall, not ones picked for this page). Its rewrites and restructures work toward items 1 to 3; check them, and everything in items 4 to 11, against the list before anything goes live.

Measure the result

Run Page Audit again after publishing to see the page score move across its six dimensions. For the answers themselves, start the fix from an Optimize page card in the Action Center. Moving the card to Done starts its measurement window, counted in tracking runs: the first reading is due after 2 runs, and a page fix gets 3 before "no effect" is a fair verdict. Each card takes one reading and keeps it, so press Re-measure now once its run counter reads 3/3. The reading covers the metric the recommendation was written for, so check the page's own questions on the Prompts page as well.