AI in Content Creation
Using AI to create content breaks nothing by itself: Google evaluates whether the content is helpful, not who typed it. The trouble starts where generation replaces substance. This lesson covers the questions Google suggests asking of your own text, where the line runs between a draft and a fake, and what a person does once the draft exists.
We work from two documents: the helpful content guidance and Google's position on AI-generated content. Both are written in plain language and are worth reading in full.
Thousands of sites repeated this story in 2026. The "what share of the top results is AI-written" and "how far traffic fell" figures circulate without a primary source, and they cannot be checked, because authorship is not detectable from the outside. Something else is checkable: Google published the list of questions worth asking of your own material. That is where we start.
Why one technology produces opposite outcomes
Two sites use the same tool and get different results. The difference is not the model but what happens after generation. Google frames it as three questions worth asking of any material: who created it, how it was created, and why.
| Question | What it tests |
|---|---|
| Who | Whether the reader can tell who wrote it. A byline, a bio, what the person does |
| How | If automation did a substantial part, whether that is self-evident to the reader and whether you explain why it was useful |
| Why | Whether the material exists to help a person or to collect visits from search. Google calls this the most important question |
The wording on "why" is blunt: if automation is used for the primary purpose of manipulating search rankings, that is a violation of the spam policies. Not "bad practice" — a violation, with everything that follows.
That is the difference between the two sites. One uses generation to reach the substantive work faster. The other uses it to avoid doing that work.
What Google suggests asking of your own text
The industry settled on the term "information gain" — whether a page adds something others do not have. The idea is right, but Google does not use the term: in the helpful content guidance the same meaning is a question — does the content provide original information, reporting, research, or analysis? And right there is another worth reading twice: are you mainly summarizing what others have to say without adding much value?
Pure generation answers that second question with "yes" by default: the model reassembles the existing web. To change the answer, a person has to add what the model does not have.
| Summarizing others | Original contribution |
|---|---|
| The top ten results, rephrased | Your own data: measurements, a customer survey, the result of an experiment |
| "Experts agree", "works best" | Specific numbers with a source, a date and who measured them |
| Text with no trace of an author or real experience | Screenshots of your own reports, customer quotes, an honest account of your own failure |
| A long definition that a reference work states better | A connection between several things that nobody drew before you |
| Universal advice with no context | Advice for a specific case: site size, niche, market |
To see how much of the material is generalities, the text quality check and the generation-signal check help — not to fool Search, but as a signal the text came out flat. The term is in the glossary: thin content.
The working process: where the machine belongs, and where a person does
The "write it yourself or generate it" argument is framed wrong. It is more useful to split the work by what each side does better. A machine assembles structure and smooth phrasing well. A person brings what is not on the web yet.
| Stage | The machine | The person |
|---|---|---|
| 1. Recon | Reads the top ten, what the topic covers and what is missing | Decides which subtopics are worth taking at all |
| 2. Structure | A draft set of headings from the audience's questions | Sets the angle: what will make this one different |
| 3. Draft | A first pass over the structure | Does not write — saves the effort for the next step |
| 4. Your contribution | Helps format the edits | The point: adds data, cases, screenshots, quotes, conclusions from actual practice |
| 5. Finishing | Meta tags, alt text, a draft of the internal links | Reads it as the audience would and fixes the tone |
How long that cycle takes depends on the topic and on whether you have data of your own: there are no norms here, and figures like "twenty minutes per article" describe nothing. Measure not the time but the share of the material that was not on the web before you.
Step 4 is the only irreplaceable one. Without it the article stays a retelling of the top ten, however smooth the prose.
How to pick a tool for the job
Rankings of "which model writes best" go stale between releases, and you cannot verify them anyway. It is more useful to hold not a list of brands but a test: what is this tool actually fit for?
| Job | What the tool has to do |
|---|---|
| Research a topic | Search the web at query time and show sources, rather than recite from memory |
| A long coherent draft | Hold the whole article in context and not lose the tone by the middle |
| Heading and snippet options | Produce many options quickly, so there is something to choose from |
| Check topic coverage | Compare your text with what competitors cover — that is the entity coverage analysis |
| Quality check before publishing | Show readability and density — the readability analysis and the density analysis |
One practical caveat: the more tools in the chain, the more places your voice leaks out. Start with one for the draft and one for the check.
How to state the task to a model
The quality of the draft depends heavily on what you asked for. How heavily cannot be measured, and "three times faster" from marketing posts has nowhere to come from. What is clear is what the model lacks when the result comes out bland: context.
📋 Four parts that are usually enough:
Role. Who the model should be and who it writes for. Input. The topic, the audience, what is already in the results, how you want to differ. Constraints. Length, number of sections, tone, what not to do. Examples. One or two paragraphs in the style you want — that works better than any description of style in words.
The last part gives the most and is used the least. "Write like an expert, no fluff" can mean anything to a model; two paragraphs of your own text mean exactly one thing.
For long pieces it is safer to split the task into steps — recon, structure, draft, finishing — where each step receives the previous one's output. That way you can see which step went wrong instead of rewriting the whole article.
What not to do
| Approach | Why it does not work |
|---|---|
| "Generate a thousand articles and publish them" | Google's policies name this specifically — scaled content abuse. The rule is not about AI: it applies the same to generation, to human production and to a mix |
| "Publish it as it comes" | An unedited draft answers "yes" to the question about summarizing others. Readers see it too: generalities, and no example that could not have been taken from the results page |
| "Replace the expert with a model" | The model does not know your customers, your numbers or your failures. Those are exactly what competitors do not have |
| "Run it through a rewriter so it is not detected" | Google evaluates helpfulness, not authorship — hiding the signs of generation solves a problem that does not exist and makes your own editing harder |
| "Hide that we used AI" | The documentation suggests the opposite: if automation did a substantial part, say so and explain why it was useful |
That last row is what the arguments about AI usually forget. Disclosure is not penalised; what is penalised is material made for rankings. If you explain to the reader how the text came about, you are answering exactly the question Google suggests asking.
Six rules worth applying today
1. Decide in advance what you will add yourself
Before opening the model, answer this: what in this piece will be mine? One measurement, one screenshot of a report, one customer quote, one account of your own mistake. If there is no answer, do not start — the retelling of the results page is already written.
2. Give examples instead of describing style
Two paragraphs of your own text in the prompt work better than a page explaining what the tone should be. It is the cheapest change that visibly alters the draft.
3. Edit substance, not style
The main editing mistake is an hour spent smoothing phrasing. The right pass: for each section, ask "what do I know about this that is not in the draft" and write it in. One checkable fact, number or example per section is a reasonable bar.
4. Write the first and last paragraph yourself
In the first, a person decides whether to keep reading. In the last, the impression settles. A model handles the middle decently; its openings and endings come out impersonal.
5. Put a name on the piece
A byline saying what the person does answers the "who" question. It works regardless of who typed the text, and it accumulates: one author writing regularly on a topic becomes recognisable to readers and to Search alike. The term is in the glossary: E-E-A-T.
6. Count what happens to the articles, not how many there are
Twenty published pieces of which three get impressions are not twenty pieces, they are three. The content decay finder tracks which pages lose traffic over time.
Practice: run your own material past Google's questions
Four steps on your own text:
- Open the helpful content guidance and copy out the five questions closest to your topic.
- Take your most recent article and answer each one honestly. Count "sort of" as "no" — it means the trait is absent.
- Find the places where a generality stands in for a fact. The text quality check catches those, and entity coverage shows what the topic is missing.
- A month later, compare that page's impressions in the performance report against its neighbours. That is your answer about usefulness — yours, not from someone else's statistics.
The glossary covers the terms: E-E-A-T, thin content, search intent, AI overviews.
How to tell you are using AI properly
The sign is simple and needs no tooling: look at where your time goes. If it goes into making the draft sound more alive, you are solving the model's problem. If it goes into finding and adding what is not on the web, you are solving the reader's.
Three questions to check yourself. Can you name what in your last article was yours? Does the reader know who wrote it? Would you have written it if search did not exist?
What not to expect: that speed alone produces results. Twenty articles a month with no contribution of your own lose to four that have one — and they lose not on position but on the fact that nobody needs the first twenty.
What comes next
That completes the "AI and search" module. Next up: the anatomy of a results page — what blocks it is assembled from and which slot you are really competing for. The previous lesson, on how language models work, is here.