AI and SEO: what changed and what did not

If you have spent an evening reading about AI and search, you have probably collected two incompatible beliefs. The first is that Google will punish you for publishing anything a model touched. The second is that Google will stop sending you traffic at all, so publishing is pointless anyway. Held together, they produce paralysis, which is a bad state to make decisions from.

Both beliefs come from the same mistake: three separate developments have been crushed into one phrase. Pull them apart and most of the anxiety has nowhere to sit.

| The thing | The actual question | Who decides the answer | |---|---|---| | AI-generated content | Will pages produced with a model be penalised? | Google's spam and quality systems | | AI-assisted workflows | Does using a model make my work worse? | You, and your editing standards | | AI-powered search results | Will generative answers take my clicks? | User behaviour, still unfolding |

The first has a documented answer. The second is a craft problem you control entirely. The third is genuinely unresolved, and anyone who tells you otherwise with a number attached is guessing in a confident voice.

We should say plainly, before going further, that we use AI-assisted production for articles on this site, including this one. Writing about the subject while implying otherwise would be a strange kind of dishonesty. What that means in practice — and where the humans sit in it — is covered further down, because it is the most useful part of the answer.

The content question has a documented answer, and it is boring

Google's documentation addresses AI-generated content directly, and the position has been consistent: the company rewards high-quality content however it is produced. Automation is not itself the problem. Using automation to generate large volumes of unhelpful content intended primarily to manipulate rankings is the problem, and Google's spam policies describe that as scaled content abuse — a category defined by purpose and quality, not by which tool typed the words.

Read that definition slowly, because the important word is *primarily*. The test is what the page is for. A page made to rank, which happens to contain some information, fails. A page made to answer somebody, which happens to have been drafted with a model, does not automatically fail anything.

This reframes the worry usefully. "Will I get penalised for using AI?" is the wrong question — you cannot answer it, because it depends on facts about your page rather than facts about your tooling. The answerable version is:

• Would a person with this problem finish this page better off?
• Does it contain anything that is not already on the first ten results?
• Would you send it to a customer who asked you the question directly?
• If you removed the search phrasing, would the page still have a reason to

exist?

If those answers are yes, the production method is not your risk. If they are no, the production method is not your risk either — you have a much older problem, and it is the same one that sank thin human-written pages long before any of this. Our piece on how Google ranking works in plain terms covers the underlying machinery, which has not been replaced by anything.

The detection worry, specifically

People ask whether Google can detect AI text. Treat the question as a distraction regardless of the answer. Google's stated standard is helpfulness and quality; a system that could perfectly identify generated text would still have no documented reason to demote a genuinely useful page for it. And "AI detector" tools sold on the promise of a percentage score are notoriously unreliable in both directions — they flag careful human prose and clear generated prose more or less alike. Optimising a page to score well on one of them is effort spent pleasing a machine nobody at Google is consulting.

The workflow question: how to use models without producing filler

This is the part you actually control, and where most of the damage happens. The risk is not penalty. The risk is producing something indistinguishable from everything else, at a speed that lets you produce a great deal of it.

Generated drafts have a characteristic failure. They are fluent, structurally tidy, and hollow — the shape of expertise without the load-bearing specifics. A model can write "consider your local competition carefully". It cannot tell you that the two firms outranking you both have a Google Business Profile with weekly photo uploads and you have neither, because it has never looked.

A worked illustration. Take Harrow Damp Solutions, an invented two-person damp-proofing firm, writing a page on rising damp. The model draft explains what rising damp is, lists causes, and recommends a professional survey. It is accurate, competent, and identical to forty other pages. What the model cannot supply is the material the business already owns: that most Victorian terraces in their patch have a slate DPC rather than none at all, that the commonest misdiagnosis they attend is condensation from a blocked airbrick, and a photograph of the airbrick. That takes the owner twenty minutes on the phone. The draft goes from filler to the only page on the subject written by somebody who has stood in those houses.

That is the whole workflow in one sentence: use the model for the parts that are generic, and spend the time you saved on the parts that cannot be.

Some practical division of labour that holds up:

Reasonable uses. First drafts of structure. Rewriting a section you have already thought through but phrased badly. Summarising your own notes, call transcripts or support tickets into themes. Producing variants of a heading so you can pick one. Compressing something long that you wrote and understand. Checking whether you have answered the question you set at the top.

Unreasonable uses. Facts. Statistics. Anything with a number, a date, a citation, or a claim about what a study found. Models produce plausible figures with total composure and no source, and a fabricated statistic in your published work is a trust problem you may never learn about, because the reader who spots it does not write to tell you. If a figure is not one you can point at, do not publish it. Say the qualitative thing instead — "most sites in this position" is honest; an invented percentage is not.

The rule that matters most. Somebody accountable reads every published word and would defend it in a room. Not skims — reads. If nobody in your organisation can be named as responsible for a page being true, you are not publishing, you are emitting.

Volume deserves its own warning. The economics of generation invite scale, and scale is precisely what Google's spam policies name. A site that publishes forty mediocre pages a month is not forty times better off than one publishing two good ones; it is likelier to have diluted whatever quality signals it had. If you are deciding what to make at all, a content strategy built from your own customers' questions will beat any volume target.

The search results question, where the honest answer is that nobody knows

Generative answers — the summarised responses appearing above and instead of links, in Google and in assistants people now use directly — change the shape of a search result page. What they do to your traffic over a period of years is not something anyone can currently tell you, and you should be suspicious of any figure presented as settled. The public data is fragmentary, the products are changing month to month, and most circulating numbers come from vendors selling a remedy.

What can be said without inventing anything:

Zero-click results are not new. Weather, conversions, dictionary definitions, sports scores, opening hours — search stopped sending clicks for whole classes of query long before generative answers arrived. What has changed is that the boundary moved outward, into explanatory queries that used to produce a visit.

Exposure is uneven, and predictably so. Queries whose entire value is a short factual lookup are the ones a generated answer resolves completely. Nobody needed your 1,400 words to learn what a standard bathroom door width is. Queries carrying a decision, a risk, a price, or a local commitment behave differently: a summary can inform them, but the person still has to choose a supplier, see the evidence, and be reassured by something specific. That is a click a paragraph cannot substitute for.

The traffic that survives is likelier to be worth more per visit. If the casual browsers are being served upstream, the remaining visitors are further along. Whether that offsets the loss in volume depends entirely on your business, and you will find out from your own analytics before you find out from an article.

A rough triage you can run on your own pages this afternoon:

| Page type | Exposure to generative answers | Sensible response | |---|---|---| | Definitional, one-fact, "what is X" | High | Keep for context, stop expecting traffic from it | | Comparison and "which should I choose" | Moderate | Add the specifics only you have — prices, trade-offs, failures | | Local service and "near me" | Lower | Proof, coverage, reviews, evidence you exist physically | | Original data, method, or first-hand account | Lowest | This is where to spend effort | | Transactional pages people arrive at ready | Low | Leave alone, fix the conversion path |

Notice that none of these responses is a new discipline. They are the things that were already sensible, with the pressure redistributed. If your pages are not ranking at all, the causes are usually much more ordinary than AI; a diagnostic checklist for a site that is not ranking is a better first stop than any of this.

What "structuring content for machine extraction" actually means

This phrase has attracted a great deal of jargon — answer engine optimisation, generative engine optimisation, a fresh acronym every quarter. Strip the packaging and there is a real, modest idea underneath: if a system is going to lift a passage out of your page, the passage has to survive being lifted.

Practically:

1. Answer the question near where you ask it. If a heading poses a question,

the paragraph beneath it should contain the answer, not a preamble that arrives at the answer eight hundred words later.

1. Write self-contained paragraphs. A passage that depends on three earlier

paragraphs to make sense cannot be extracted intact. This is also just better writing for humans who scan.

1. Use headings that describe content, not headings that are clever. The

same discipline that helps a reader skim helps anything else parse.

1. Put comparable facts in tables and lists. Structure that is visible in

the markup is structure that survives.

1. Add structured data where it genuinely describes the page — a product, a

recipe, an event, an organisation. Google's documentation is clear that structured data helps a page qualify for particular result features; it is not a ranking lever, and marking up things you do not have does not create them.

1. Keep URLs stable and pages crawlable. None of the above matters if the

page cannot be fetched. The unglamorous foundations are covered in our technical SEO audit framework.

What this is not: a separate optimisation channel you can buy. If somebody offers to optimise your site for AI answers, ask how they will measure the result. If the measurement is a screenshot of one chatbot mentioning the brand once, that is not measurement, it is a demonstration.

Being cited by a generative answer also tends to depend on being visible and credible in the ordinary way — the same corroboration and reference that underpins how backlinks actually work. It is not a shortcut around reputation. It runs through it.

Common mistakes we see people make right now

• Stopping publishing entirely because traffic might fall. The alternative

to publishing something useful is not safety; it is absence.

• Publishing far more because generation is cheap. The cost that mattered

was never the typing.

• Deleting pages that lost impressions without checking whether they still

convert. Impressions and enquiries are different columns.

• Buying an "AI visibility" package from a vendor who cannot state a

measurement method before taking money.

• Rewriting an entire site to sound less like a machine. If it is accurate,

specific and useful, its cadence is not the problem.

• Expecting a faster timeline because the drafting got faster. Indexing,

evaluation and trust move at their own pace; what the evidence supports about SEO timelines has not been rewritten by better tooling.

Where a paid listing fits, and where it does not

We run a marketplace where businesses pay to hold a public leaderboard position, so treat what follows with the appropriate suspicion and check it against our published methodology.

The narrow honest claim: a marketplace position is a distribution channel whose mechanics do not depend on a search algorithm or on being cited by a model. The public rankings and what each position costs are visible to anyone, which means you can evaluate it before spending rather than after.

The equally honest limits. It does not affect your Google rankings — that is stated plainly in what pay-to-rank cannot do. It does not make you appear in generative answers. It sends a modest, verifiable amount of attention to a business that must still have something worth arriving at. If your underlying problem is that your site has nothing on it a customer needs, a listing relocates that problem rather than solving it, and the organic work described above is the better use of the same money.

Questions people actually ask

Will Google penalise my AI-assisted article? Not for the method. Google's documentation focuses on whether content is helpful and original; the spam policies target content produced at scale primarily to manipulate rankings, however it was made.

Should I disclose that AI was involved? There is no ranking requirement to. Do it where a reader would reasonably want to know — as we have here — and always where accuracy is at stake, such as medical, legal or financial material.

Is SEO dead? No, and be careful who says it is. The mechanics of crawling, indexing, relevance and reputation still determine what appears. What changed is the value of certain query types, unevenly. If you are weighing whether to hire help, a decision framework for hiring an SEO agency is more useful than a verdict on the industry.

How much traffic will I lose to generative answers? Unknown, for you specifically and for everyone in general. Check your own analytics by query type over the coming quarters. That measurement is real; the industry-wide figures being quoted at you mostly are not.

What should I do first, this week? Run the triage table above over your top twenty pages. Find the ones that are pure lookup and stop investing in them. Find the two where you know something nobody else has written down, and write it down.