Using AI in content production without losing quality or rankings

Language models changed content production quickly. A draft that took a writer a day now takes a minute, and many sites responded by publishing far more than before. Some of that output is fine. A lot of it is fluent, correctly structured and empty: pages that restate what already ranks, contain no first-hand information, and occasionally state things that are not true. Search engines have spent the last few years getting better at recognising exactly that kind of page, and readers recognise it even faster.

This guide is about using AI where it makes content better or cheaper without making it worse, and about the workflow and checks that keep quality up.

What Google has said

Google's public position has two parts, and both matter:

• How content is produced is not the issue. Google has said its systems aim to reward helpful, original content that demonstrates experience and expertise, whether it was written by people, with AI assistance, or both.
• Scaled, unhelpful content is. Its spam policies describe "scaled content abuse": generating many pages primarily to manipulate rankings with little value for users. The policy explicitly covers content created with automation, including generative AI, as well as by people or templates.

Google also suggests that where readers would reasonably wonder how content was made, explaining the process (who wrote it, how AI was involved, how it was checked) is good practice.

The practical reading: AI-assisted pages compete on the same terms as any other page. The difference is that AI makes it very easy to produce large numbers of pages that fail those terms.

Where AI genuinely helps

Task	Why it works	What to watch
Summarising your own material	Interview transcripts, support tickets, product notes become usable drafts quickly	Check nothing was added that was not in the source
Generating question lists	Surfaces follow-up questions readers might have	Validate against real data: PAA, support logs, Search Console
Outlining from a brief	Fast first structure to react to	Outlines drift toward the generic average; edit hard
Editing and tightening	Shorter sentences, clearer structure, consistent style	Can flatten voice and remove useful specifics
Formatting	Turning prose into tables, lists, FAQs	Check numbers survived intact
Translation drafts	Good starting point for localisation	Needs native review; see localizing keywords
Data work	Classifying queries, grouping keywords, writing regex	Spot-check classifications

Where AI quietly lowers quality

• Writing the substance. A model knows what is commonly written about a topic. It does not know what you found when you tested the product, what your customers keep asking, or what your data shows. Articles written from the model's knowledge are, by construction, averages of what exists.
• Facts, numbers and citations. Models produce plausible statistics, dates and quotations that are sometimes wrong and sometimes invented. Every figure in an AI-assisted draft needs a source you checked yourself.
• Specialist and YMYL topics. Health, finance, legal and safety content needs qualified human authorship and review; see E-E-A-T in practice.
• Volume for its own sake. The cost of publishing fell; the cost of being judged by search engines and readers did not. A hundred mediocre articles can weigh on a site more than they help.
• Voice. AI drafts converge on the same tone and phrasing. Readers notice, and brands lose the distinctiveness that made people remember them.

A workflow that holds up

1. Start from a human brief with a defined reader, angle and "what we add" section. Without it, AI produces the average. See content briefs.
2. Gather the original material first: test notes, photos, interview transcripts, data. This is the part only you can supply.
3. Draft, with or without AI, from that material. If AI drafts, give it your notes and instruct it to use only them for facts.
4. Edit as a subject expert: remove generic padding, add specifics, correct anything vague or wrong, restore the voice.
5. Fact-check every claim against a source you opened. Delete anything you cannot verify.
6. Read it as the target reader: does it answer their question better than the current top results? If not, it is not finished.
7. Publish at a pace you can edit. If editing is the bottleneck, publish fewer pages, not less-edited ones.

One article, two workflows

A hypothetical kitchenware shop wants an article on choosing a cast iron skillet. In the first workflow, someone asks an AI tool for "a 1,500-word SEO article on the best cast iron skillet", lightly edits it and publishes. The result covers pre-seasoned versus bare, weight, handle length and care, in the same order as dozens of existing articles. It recommends "considering your cooking needs". It mentions a weight figure that is wrong for the shop's own best-seller. Nothing in it could only have come from this shop.

In the second workflow, the brief asks the shop's buyer, who has cooked with every skillet they stock, for notes: which pans customers return and why, which handles get too hot, how the weights actually compare on the shop's kitchen scale, what the seasoning looks like after a month of daily use. Someone records a twenty-minute conversation. AI turns the transcript into a structured draft using only that material, an editor tightens it and adds photos from the shop's test kitchen, and every number is checked against the scale and the product sheets.

Both took a fraction of the time an article used to take. Only the second one gives a reader, a search engine or an AI assistant a reason to prefer it.

Signs a page reads as machine-made

Editors learn to spot these quickly, and so do readers:

• An introduction that restates the title and promises to "explore" the topic.
• Every section the same length, every list exactly three or five items.
• Advice that applies to any business in any industry.
• No numbers, names, places, prices or dates specific to the subject.
• Conclusions that summarise the article instead of saying anything new.
• Stock phrases that recur across pages on the same site.

None of these prove anything about how a page was made, and human writers produce them too. They are signs that a page lacks specifics, which is the actual problem. When an editor finds them, the fix is rarely rewording; it is going back to the source material, or to the person who knows the subject, and adding what the page is missing.

If you already published at scale

Many sites now have hundreds of AI-drafted articles published with light editing. Treat them like any other content inventory. Find which pages earn traffic, links or conversions, and which overlap or say nothing new. Improve the useful ones with original material, merge overlapping ones, and remove the rest. The pruning guide describes the process. It is less painful than it sounds, because the weakest pages usually bring almost no traffic anyway.

AI content and AI search

There is an irony worth noting. AI assistants and AI-generated search summaries look for sources that add something: specific facts, clear statements, original data, first-hand detail. Pages that are themselves averages of existing content have little for them to cite. The same qualities that make a page worth ranking make it worth quoting; see writing content AI assistants quote. You can check whether assistants mention your brand for the questions that matter with the method in the AI visibility guide.

A policy you can write in an afternoon

Teams work better with a short written policy. A workable one answers: which tasks AI may be used for; which topics require a named expert author or reviewer; that every factual claim must be checked against a source; who signs off before publishing; and how AI involvement is disclosed, if at all. Review it every six months as the tools and the search landscape change.