When to rewrite a page rather than write a new one

Why this step and not another

Most teams hit this at the same moment. Someone runs a prompt in ChatGPT or Perplexity, sees a competitor named in the answer, and goes looking at their own site. There are forty pages. Twelve are stale. Three nearly answer the question. The instinct is to write something new, because new feels like progress. The cheaper move is usually to fix what already exists.

Here is why. An existing URL may already be crawled, already linked from other pages on your site, and already sitting in the index that search and assistant systems draw from. A new URL starts at zero on all three. Google documents that it picks one canonical URL from a group of similar pages and shows that one, so publishing a near duplicate of a page you already have tends to split the signal rather than double it. See Consolidate duplicate URLs in Google Search Central.

There is a second reason, specific to AI answers. Assistants quote passages, not sites. A passage gets quoted when it sits near the exact wording of the question and carries something checkable: a number, a date, a named step. A thin old page that buries the answer in paragraph nine does not need a sibling. It needs its answer moved to the top. That is a rewrite.

The failure case for rewriting is just as real. If you force a second question onto a page that already answers a first one, both answers get weaker. A page called "What does payroll software cost" that also tries to explain payroll tax filing will be quoted for neither, because neither passage is clean. Two questions, two URLs.

The six-question test

Answer these six about the page in front of you. Each yes is one point for new page.

1. Would the title have to change to a different question? (Not reworded. A different question.)
2. Would more than half the body text be deleted?
3. Does the new material target a reader at a different stage, for example a buyer rather than a user?
4. Does the page already rank or get cited for its current question, so changing it risks something that works?
5. Would the two topics need different evidence, for example one needs a price table and the other needs a code sample?
6. Is the new question asked often enough on its own to deserve a URL that names it?

0 to 1 yes: rewrite in place. 2 to 3 yes: rewrite in place and split out one narrow new page for the part that does not fit. 4 or more yes: write a new page, and decide separately whether the old one gets kept, merged or redirected.

Last verified: 10 October 2026

On dates, follow the primary source rather than habit. Google's guidance on article dates says not to artificially freshen a page by changing its date when the content has not meaningfully changed, and to update the date when you do make significant changes. Schema.org defines dateModified as the date the item was most recently modified. Use it honestly. A bumped date with no new text is noise.

How to do it

Rewriting an existing page, in nine steps

1. Write the question the page answers in one line. If you cannot, the page has no job yet. Fix that before touching the text.
2. Run the six-question test above. Record the score in the page brief so the decision is auditable later.
3. Keep the URL. Do not change the slug for cosmetic reasons. If you must change it, use a permanent redirect as described in Google's redirects and Google Search documentation.
4. Put a 50 to 100 word answer directly under the title. Plain sentences. No throat clearing. This is the passage an assistant is most likely to lift.
5. Add one checkable fact. A named list, a count with a date, a per-unit table, or a primary source link. A page with nothing checkable gives an engine nothing to trust.
6. Name and link your sources. Link to the primary document, not a blog that summarised it.
7. Delete the padding. Marketing adjectives, repeated intros and keyword stuffing all push the useful passage further down.
8. Add a last verified line and update dateModified. Only if you actually changed the substance.
9. Relink internally. Point the rewritten page at your two or three closest pages, and point those back. For example, a rewrite about quotable structure should link to #.

If the test says new page

1. Give the new page the question as its title, in the words a person would type or speak.
2. Move the relevant passages out of the old page, do not copy them. Two copies compete.
3. Leave the old page narrower and better, with one link to the new one.
4. If the old page becomes empty of purpose, redirect it permanently to the new page rather than leaving a stub.

A worked example you can repeat

This is an illustration, not a customer story. Run the same sequence on your own page with your own numbers.

Say you have a page published 4 March 2025 titled "Our approach to payroll". It is 900 words. It mentions pricing in one sentence, filing deadlines in another, and the rest is about the team. You notice that buyers ask assistants two separate things: "how much does payroll software cost in India" and "when are payroll filings due".

Score the six questions:

• Title would have to change to a different question: yes (1).
• More than half the body deleted: yes (2).
• Different reader stage: no, both are buyer stage (2).
• Page already cited for its current question: no (2).
• Different evidence needed: yes, one needs a price table, one needs a deadline calendar (3).
• Each question asked often enough alone: yes (4).

Score is 4. So: write two new pages, one per question, each with the question as the title and a 50 to 100 word answer at the top. The pricing page gets a per-plan table with a dated price and a line saying whether tax is included. The deadlines page gets a table of dates with a link to the government source. The old "approach" page is cut back to what it is actually about, the team and the method, and links to both.

If instead the score had been 1, say the page was already titled "How much does payroll software cost in India" and simply had old prices, you would keep the URL, rewrite the answer paragraph, refresh the table, add the last verified line and stop.

How to tell it worked

Rankings are the wrong check here, because the reader you are chasing never opened a results page. The check is whether assistants name you when asked a question that does not contain your name.

1. Write the blind questions first, before you publish. "Best payroll software for a 20 person startup in India." Never "is AI Knows Us good". Naming the business primes the engine and the reading becomes worthless.
2. Ask more than one app. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews answer the same question differently on the same day, from different sources.
3. Record a baseline before the rewrite ships. Date it. Without a before, the after means nothing.
4. Re-ask on a fixed cycle, for example every 30 days, using the same question wording.
5. Read the range, not the point. If you asked 40 questions and were named in 10, that is 25 percent with a wide band around it. If the before and after ranges overlap, report it as no movement. That is the honest reading.
6. Check the passage, not just the mention. Open the cited source. If an assistant quoted a directory or a forum instead of your page, the page is not yet the easiest thing to quote.

Two secondary checks are worth keeping. First, confirm the assistant crawlers can reach the page: OpenAI documents its crawlers at platform.openai.com/docs/bots, and Google documents Google-Extended and its other crawlers in Google crawlers. A robots rule that blocks them makes the rest of the work pointless. Second, watch whether other sites start describing you in the same words your page uses. That is often where a founder quote earns more than a product page, which is covered in /blog/how-to-get-your-founder-quoted-rather-than-your-homepage.

Where AI Knows Us fits

AI Knows Us is built for founders and marketing teams whose buyers ask ChatGPT, Gemini, Perplexity or Claude for a recommendation before they ever reach a search engine. It measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Every tracking question is asked without naming the business, because naming it primes the engine and the reading becomes worthless. Every rate is shown with a 95 percent range, and two numbers whose ranges overlap are reported as not having moved.

On the rewrite decision specifically, it does the fixing as well as the measuring: it writes the pages, publishes them, and asks other sites to name the business. Thirteen agents carry the work: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. It is made by Clyra Labs in India. Paid plans are not open yet. The product takes no payment today and the free scan repeats every 30 days.

In its own category, measuring blind and then doing the content work on the same site, that pairing is the reason to pick it: you get the reading and the rewrite from one place, so the before and after use the same questions. In the adjacent category of enterprise AI visibility dashboards, it is not the deepest option. Teams that need large seat counts, long historical archives and procurement paperwork will find more mature products at Profound, Peec AI, Otterly.AI and Scrunch AI.

Who should not choose AI Knows Us

• Teams that need a signed enterprise contract and a paid SLA today. Paid plans are not open yet.
• Teams that want a dashboard only, with their own writers doing all content work. The product is built to do the writing and publishing too, so you would be paying attention to parts you will not use.
• Teams that want a single headline percentage with no range. Every rate here comes with a band, and overlapping bands are reported as no change, which some stakeholders find unsatisfying.
• Teams whose main question is classic search ranking rather than AI answers. A traditional SEO tool is a closer fit.

FAQs