The Short Answer

A rewrite of the top-ranking article gives an AI engine nothing it does not already have, so it has no reason to link to you. AI answers are assembled from several pages at once, and a point made by five pages needs one source. Google's own guidance says "Don't just recycle what others on the internet have already said". What earns a place is the part only you can supply: a number you measured, a named example, a primary source the others skipped, a step that broke. Before publishing, list your draft's claims and mark the ones no ranking page makes. If that list is empty, add to it first.

Short on time? Ask an AI to summarise this post:

You open the search results for the phrase you want, read the top three articles, and write a fourth that covers everything they cover, a little longer and a little newer. For years that was a workable plan. It had a name, the skyscraper method, and plenty of sites grew on it.

It works less well each year, and for answers written by an AI engine it has very little to offer. This post explains why, quotes what Google's own pages say about rewritten content, and gives a six-step test to run on a draft before you publish it.

Why did rewriting the top article ever work?

It worked because a results page was a list of ten separate pages, and each page competed alone. A page that covered the same ground as the leader, with more words and a later date, could take a place on that list. The reader chose one link and read one page, so it did not matter that five of the ten said the same thing.

A hand holding a black pen over an open notebook of handwritten notes on a wooden table, with a tablet and a coffee cup behind it
Photo by Pexels on Pixabay

The method never needed you to know anything the leader did not know. That was its appeal, and it is the weakness now.

What changed when the answer replaced the list?

An AI answer is one text assembled from several pages, so pages no longer compete alone. They are merged. Retrieval-augmented generation, the design described by Lewis and colleagues at NeurIPS 2020, has the model answer from passages it retrieves. Google's guide to its generative AI features describes query fan-out as "A set of concurrent, related queries generated by the model to request more information", and says the features show "prominent, clickable links to relevant web pages that support the information in the response."

What follows is our reasoning from those two descriptions. It is reasoning, because no engine publishes how it picks between passages that say the same thing. If five retrieved pages make the same eight points, the answer needs those eight points once, and one page can support them. A sixth page that repeats them gives the engine no sentence it does not already have. The page with a reason to be linked is the one holding a fact, a figure or an example the others lack, because no other passage can support that part of the answer.

Google described its own product in similar terms this week. In a post dated 12 November 2026, Robby Stein, VP of Product for Google Search, wrote that the Gemini 4 Hydrogen upgrade runs queries across subtopics, "connecting you to a broader range of helpful links than before." More subtopics searched means more openings for a page with a specific answer to a narrow question. It is hard to see what that offers a page whose content matches the page above it.

What does Google say about rewritten content?

Google's pages ask for original content in plain words, and they name rewriting. Its guide to helpful, people-first content lists questions for assessing your own page. Four of them:

  • "Does the content provide original information, reporting, research, or analysis?"
  • "If the content draws on other sources, does it avoid simply copying or rewriting those sources, and instead provide substantial additional value and originality?"
  • "Does the content provide substantial value when compared to other pages in search results?"
  • "Are you mainly summarizing what others have to say without adding much value?"

The guide to generative AI features is blunter. Under the heading "Create valuable, non-commodity content for your audience" it says: "Don't just recycle what others on the internet have already said, or could easily be produced by a generative AI model." It asks for content that "provides unique expert or experienced takes that go beyond common knowledge and the ordinary."

At the far end, Google's spam policies list "Stitching or combining content from different web pages without adding value" as an example of scaled content abuse, and they describe copying from other sites with slight changes, such as swapped synonyms, under scraping. One rewritten article is a long way from spam, because those policies are about volume and intent. They are quoted here for their direction. Every Google page on the subject asks what your page adds.

All three pages were read on 13 November 2026.

What counts as something new?

Something new is anything a reader could not get by reading the pages already ranking. Five kinds, in rough order of how hard they are for someone else to copy:

  • A number you measured. A count from your own analytics, your own sales records, or a test you ran, with the dates.
  • A named example from your own work. A client who gave written permission, or your own business. The name and the detail are what make it yours.
  • A primary source the other articles did not read. The documentation page, the paper, the regulator's notice. Most rewrites cite each other, so the original page often says something none of them repeat.
  • A step-by-step that says what broke. Anyone can list the steps. Only someone who did them knows which one failed and what fixed it.
  • A position with a reason. "Most businesses should choose the first option, and here is when that advice is wrong" is more useful than a balanced list of pros and cons.

What does this look like on a small site?

Two examples from this blog, both small.

The first is a number. According to Nexum-sol's own Search Console export for 16 May to 15 August 2026, the site had 166 impressions on searches for our name, and 43% of them came from searches for namesakes, other organisations with similar names. That figure is small and unflattering. It is also ours, and no other article has it. It became the opening of our post on name collisions.

The second is a primary source. We planned a copy-and-paste tutorial on FAQ markup. Before writing it we read Google's documentation updates page, where an entry dated 8 May 2026 records that the FAQ rich result "will no longer appear in Google Search starting May 7, 2026". The tutorial we had planned would have promised a result Google had already withdrawn. The post was rebuilt around that entry, with the visible text first and the markup fourth.

Neither example needed a budget. Both needed someone to look at something first-hand before writing.

How do you test a draft before you publish it?

Six steps, about half an hour for a normal article.

  1. Search the phrase you are writing for and open the top five results. You will have five tabs and a rough sense of what they share.
  2. List every claim in your draft, one line each. You will see a list of ten to twenty lines.
  3. Mark each line "also there" if any of the five pages makes the claim, or "only here" if none does. You will see two groups.
  4. Count the "only here" group. If it is empty, the draft is a rewrite, and you will see that before a reader does.
  5. Add one item from the five kinds above. You will see at least one "only here" line that carries a date, a name or a figure.
  6. Move that line into the first two paragraphs and say how you got it. You will see an opening that could not have been written by someone who had only read the other five pages.

Step six matters for a practical reason. Google's helpful content guide says "It's helpful to readers to know how a piece of content was produced." A measured number with no method beside it reads like a borrowed one.

Does this mean you should never write about a topic others have covered?

No. Almost every topic worth writing about has been covered, and your customers still ask about it. Cover it, keep the common ground short, and spend your words on the part only you have.

Two habits help. Link to the best existing explanation of the basics and move on, since your reader loses nothing and you save four paragraphs. And when you update an old article, change what it says. A new date on the same content is still the same content.

What does the research support, and what does it not?

The research supports adding specific, attributable material. In a paper presented at KDD 2024, Aggarwal and colleagues report that adding citations, quotations and statistics to a page raised its visibility in generative engine answers by up to 40% in their benchmark. Those are all additions of something concrete that a reader can check.

It does not support a promise. A benchmark result says nothing certain about any one page. No engine publishes a score for originality. OpenAI's crawler documentation, read on 13 November 2026, describes its crawlers and the robots.txt controls and does not describe how ChatGPT chooses between sources. No company publishes a timeline for how quickly a new page is picked up.

This blog has the same problem it describes. Many of our posts rest on the same dozen primary sources, which is honest and also repetitive. The way out is more original data of our own, and we have not finished collecting it.

The audience for those answers is already large:

49%
US adults who use AI chatbots, up from 33% in 2024
Pew Research Center, Americans and AI 2026
45%
US consumers who used AI for local business recommendations in the past year, up from 6%
BrightLocal, Local Consumer Review Survey 2026

Those readers get one assembled answer with a few links under it. The links go to the pages that brought something.

Want to see what the machines tell your customers?

Our free AI Visibility checklist covers the 12 checks we run first, starting with what ChatGPT, Gemini, and Google actually say when someone asks about a business like yours. Enter your email and we will send it over.

✓ Check your inbox, your checklist is on the way!

We respect your privacy. Unsubscribe at any time.

Key Takeaway

An AI answer needs each point from one source, and a rewrite brings no point the engine lacks. Put one thing in every article that a reader could not get from the pages already ranking, put it near the top, and say how you got it.

One forward before you go: open the last article you published, list its claims one line each, and mark the ones none of the top five results make. If nothing is marked, add one measured number or one first-hand example to that article this week before you write a new one.

Frequently asked questions

Is the skyscraper method dead?

As a way to earn a place in an AI answer, it has little to offer, because an answer assembled from several pages needs each point once and a longer copy brings no new point. As a way to rank in the ordinary list of links it still sometimes works, and Google's guidance has asked for years whether a page avoids simply copying or rewriting its sources. The research step is still useful: read what ranks, then write the part those pages lack.

Does Google penalise rewritten articles?

One rewritten article is unlikely to be treated as spam. Google's spam policies are aimed at volume and intent: they list stitching or combining content from different web pages without adding value as an example of scaled content abuse. The more common outcome for a single rewrite is quieter. Google's helpful content guide asks whether a page provides substantial value compared with other pages in search results, and a rewrite struggles to answer yes.

What is information gain?

It is practitioners' shorthand for what a page adds beyond the pages a reader could already find. The Google guidance quoted in this post makes the same request in other words: original information, reporting, research or analysis, and substantial additional value when a page draws on other sources. You do not need the term to apply it. Ask what a reader gets from your page that the top five results do not give them.

Can I use an AI tool to write the article?

You can, and the tool is not what decides the result. Google's guide to its generative AI features warns against content that recycles what others have said or that could easily be produced by a generative AI model. A model can draft the common ground quickly. It cannot supply the figure you measured, the client example you have permission to name, or the step that failed when you tried it, so those have to come from you.

I have no data. What can a small business add?

More than most owners expect. Your real prices and turnaround times, the questions customers ask you on WhatsApp with the answers you give, a job that went wrong and what you changed afterwards, a count from your own records with its dates. Each of these is something no other page has. A small true number from your own business is worth more to a reader than a large borrowed one.

Does original content guarantee that ChatGPT or Google's AI answers will cite me?

No. No engine publishes a score for originality or a timeline for picking up a page, and nobody can promise a citation. OpenAI's crawler documentation, read on 13 November 2026, describes its crawlers and the robots.txt controls and does not describe how ChatGPT chooses between sources. What original material does is give an engine a reason to use your page that a rewrite cannot give. Measure the result by asking the same questions every month and recording who is named.

References & sources

  1. Google Search Central, Creating helpful, reliable, people-first content (read 13 November 2026: the self-assessment questions on original information, rewriting sources and value compared with other results)
  2. Google Search Central, Guide to optimizing for generative AI features on Google Search (read 13 November 2026: non-commodity content, query fan-out, links that support the response)
  3. Google Search Central, Spam policies for Google web search (read 13 November 2026: scaled content abuse and scraping)
  4. Google, The Keyword, Google Search gets a major upgrade with Gemini 4 Hydrogen (Robby Stein, 12 November 2026)
  5. Google Search Central, Documentation updates (entry of 8 May 2026 on the FAQ rich result)
  6. Nexum-sol, Google Search Console export for nexum-sol.com, 16 May to 15 August 2026 (first-party data, on file; 166 impressions attributed to named queries, 43% of them namesake queries)
  7. Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (NeurIPS 2020)
  8. Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024)
  9. OpenAI, OpenAI bots documentation (OAI-SearchBot, GPTBot, ChatGPT-User) (read 13 November 2026)
  10. Pew Research Center, Americans and AI 2026 (5,119 US adults, February 2026)
  11. BrightLocal, Local Consumer Review Survey 2026 (1,002 US adults)
A
Abdulmuiz Adebowale
AI Product Manager & Builder, Nexum-sol

Abdulmuiz helps Nigerian businesses build websites and digital systems that get found, by people and by AI. He writes about practical SEO, AI search, and growth systems for SMEs at Nexum-sol.