Before touching the copy, read the answer as if it were a witness statement. The omissions, phrasing, and source choices usually tell you which page claim failed under pressure.
The search result was only a few lines long, but it had already made a decision. In a composite scenario from a Dublin corporate training provider, forty-six staff, a mix of HR teams, public-sector departments, and mid-sized technology clients, the AI-style answer described the category neatly. It mentioned delivery formats, compliance needs, and trainer experience. Then it named competitors whose pages were not prettier. They were simply easier to summarise.
The provider’s own site used “flexible learning” so often that the phrase had gone smooth from handling. It appeared on programme pages, the home page, and a few case-study fragments. But the answer surface did not reward the phrase. It pulled from pages that said who the training was for, how cohorts worked, what constraints applied, and what evidence supported the delivery model. The uncomfortable part was this: the provider knew all of that. The website had not made it easy to read.
The answer surface is evidence, not an oracle
When an owner sees a competitor inside an AI Overview or an SGE-style answer, the first reaction is often to rewrite the page in a hurry. I understand the impulse. The visible result feels like a verdict. Someone else has been chosen; your page has been passed over; the cursor starts blinking inside the CMS before anyone has really read what happened.
That is too quick.
An AI Overview is not a perfect explanation of why a page won. It is a visible compression of what the system was willing to say for that query. The wording, the included sources, the missing distinctions, and the order of ideas are all useful evidence. They are not holy. They can be unstable. They can omit good businesses and include thin ones. Still, before rewriting, I want to know what the answer surface appears to value.
In practice, that means slowing down and reading the result as a field object. What service category does it think the query belongs to? Does it describe a buyer situation or only a definition? Does it mention location? Does it cite process, credentials, pricing shape, constraints, reviews, or examples? Which phrases appear across the cited pages? Which claims are absent even though the owner thinks they are central?
An AI Overview SEO audit is a page diagnosis that begins with the visible answer surface, because the surface shows which claims survived compression and which proof the system could use.
That definition matters because it keeps the work from turning into guesswork with nicer headings. The audit does not begin with a keyword list. It begins with a particular answer, for a particular query, in a particular service category, then walks backward to the page that should have been understood.
The surface may be wrong. Fine. We still learn from its wrongness.
Read the shape before reading the sources
In the Dublin training-provider composite, the query was about finding corporate compliance training in Ireland. A simplified version of the answer grouped the category around delivery format, sector suitability, trainer expertise, and documented outcomes. It did not dwell on broad promises about engaging learning. It did not reward pages that merely said “tailored” or “flexible”. It preferred pages that made comparison easier.
That is the first thing to record: the shape of the answer. I do this before opening the cited pages. Otherwise the eye rushes to competitor copy and starts copying surface features. The answer’s structure is the cleaner clue. If the answer sorts providers by delivery format, your audit needs to inspect whether your page makes delivery format explicit. If it mentions regulated sectors, your page has to show whether the business serves them and what proof supports that claim. If it uses location language, your page needs local evidence rather than a county name sprinkled into the footer.
I keep a handwritten ledger of these patterns because writing them down slows the temptation to over-read. The line might be plain: “AI answer frames training choice around format + compliance + audience.” That is enough to start. Then I look at the client page and ask whether those same elements are visible above the fold, in the service body, and near proof.
This produces a different kind of rewrite brief. Instead of “make the page better for AI”, which is almost useless, the brief becomes: explain programme formats earlier; separate HR-team training from public-sector compliance contexts; state cohort size ranges if they genuinely exist; move trainer credentials into plain wording; connect a review or case fragment to the delivery promise.
Notice the restraint. The answer surface gives direction. It does not give permission to invent.
The compression gap is usually small and costly
I use the term compression gap for the distance between what a business knows about its service and what a search system can safely summarise from the page. The gap is often small. A missing noun. A buried process. A proof point trapped in a PDF. A phrase repeated so often that it hides the useful detail underneath.
In the training provider example, “flexible learning” was not false. The business did offer several formats. The problem was that flexibility had swallowed the specifics. A human salesperson could explain the difference between a live workshop, a blended programme, and a compliance-led session for a public-sector department. The website left those distinctions scattered. The answer surface, reading without the salesperson, chose pages that did the sorting themselves.
This is the part owners sometimes dislike. They want the audit to discover a technical defect. Sometimes there is one. More often, the first defect is linguistic and structural. The page has not made the business repeatable.
A page becomes repeatable when a stranger can say, in one sentence, what the service is, who it fits, when it is used, and why the claim is credible. AI systems appear to favour repeatable material because repeatable material can be compressed with lower risk. That is my observation from audits, not a universal rule. But it has held often enough that I treat it as a practical test.
The compression gap has several forms. In my notes, I mark four types: claim blur, proof distance, structure tangle, and local thinness. Claim blur means the page never states the offer sharply. Proof distance means the evidence exists but sits too far from the claim. Structure tangle means several intents are forced into one page without clear separation. Local thinness means the business names Ireland, Dublin, Galway, Cork, or another place without showing real service presence, constraints, or situations there.
A single page can have all four. That is when a rewrite becomes tempting and dangerous. If you rewrite everything at once, you may lose track of what the answer surface actually showed.
Competitor pages are clues, not templates
After reading the answer shape, I open the cited pages. Slowly. I am not looking for lines to imitate. I am looking for evidence patterns. What did these pages make explicit that the client page left implied? Where did they place proof? Did they define the service category in ordinary language? Did they separate audiences? Did they name constraints? Did they show credentials in a way a non-specialist could understand?
In the Dublin composite, one competitor page had a small roughness I liked. It was not elegant. It had a clunky section explaining which training sessions could be delivered online and which ones worked better in person. The copy would not win an award. But for the query, it did real work. Another page named cohort size bands and gave a plain account of trainer background. A third page had a slightly dated case example but tied it clearly to an HR-team problem. The answer surface seemed to prefer that usable mess over smoother language.
That is not strange. Smoothness is often overrated in service pages. A buyer comparing options does not need literary mist. They need enough firm edges to understand fit. AI-search systems seem to need the same thing, though for a different reason: they must summarise without making up the missing middle.
The risk is that an owner sees a cited competitor and copies the wrong feature. They notice a heading, a word count, a FAQ, a layout block, or a phrase. But the useful feature may be deeper. The competitor may have made the client type clear. Or pulled credentials near the claim. Or explained a process step that everyone else assumed. Or admitted a limitation that made the rest of the page more believable.
I sometimes write two columns during this stage. One says “visible in answer”. The other says “visible on cited page”. When the same idea appears in both, I mark it. Those marks become the audit’s evidence. Not proof in the scientific sense. Useful evidence for a page decision.
The aim is to understand why the page was compressible.
Rewrite only the part the audit has earned
The most expensive rewrite is the one that starts before the diagnosis is finished. I do not mean expensive only in fees. It is expensive because it smears the evidence. The page changes everywhere, rankings and answer surfaces shift for reasons nobody can isolate, and the business is left with a nicer page but no sharper understanding.
An AI Overview audit should earn each rewrite instruction. If the answer surface frames the category around delivery format, then rewrite the delivery-format section first. If the cited pages make trainer credentials plain, then bring credentials closer to the relevant service claim. If the answer mentions public-sector suitability and the client genuinely has that experience, then state the conditions and proof. If the page cannot prove the claim, leave it out until the business can.
In the training-provider composite, I would not begin by rewriting the whole home page. I would begin with the programme pages most directly tied to the query. One page may need a first-screen sentence that says the programme type, buyer, delivery format, and evidence. Another may need a comparison section separating workshops, blended delivery, and compliance-led sessions. A third may need trainer credentials translated out of internal shorthand. The work is modest in appearance. That is not a weakness.
The owner should be able to explain every change aloud. “We moved this because the answer surface sorted providers by delivery format.” “We added this because our page said flexible but did not explain the actual options.” “We removed this claim because we had no proof near it.” If the owner cannot say why a change was made, the audit has become decoration.
This is also where I am cautious with forecasts. If the current pattern holds, pages that state service fit, proof, and constraints clearly should have a better chance of being used in AI-style answers. But no honest adviser can promise inclusion from a rewrite. The output we control is page clarity. The search surface is observed, not commanded.
That distinction keeps the work sane.
Keep a record of what the answer taught you
A single AI Overview is a thin object. It should not be asked to carry a whole strategy by itself. I like to compare several queries around the same service category: broad, local, problem-led, buyer-led, and comparison-led. Then I look for repeated answer patterns. If three different surfaces reward pages that name delivery format, that is stronger than one surface doing it once. If local evidence appears only for certain query shapes, that is worth recording. If credentials matter for one query and process detail matters for another, the architecture may need both in different places.
This is where the private ledger earns its keep. I write the answer pattern by hand after an audit because the page has usually become too familiar on screen. The hand notices different things. A repeated noun. A missing location. A source that wins with plain structure. A client claim that sounds fine until the answer surface ignores it.
The record should be simple enough to revisit. Query. Answer shape. Cited-page pattern. Client-page gap. Recommended page action. Evidence needed. That is the whole skeleton. No theatre.
For the Dublin provider, the record might say: “Compliance training query framed around format, audience, trainer evidence, sector fit. Client page says flexible but does not sort formats. Proof exists in case fragments and trainer bios. Rewrite programme introduction and delivery section before adding new articles.” That note is not glamorous. It is useful.
Over time, these records show whether the site’s problem is mostly claim blur, proof distance, structure tangle, or local thinness. Then rewriting becomes less reactive. The business stops chasing every visible answer and starts strengthening the parts of the site that repeatedly fail under compression.
That is the point of reading first. Not passivity. Better aim.
The page should answer the pressure placed on it
An AI Overview puts pressure on a page. It asks the page to become shorter, clearer, and safer to quote than the page may have been written to be. Many service pages are written for persuasion, which is understandable. They want to sound confident. They want the reader to feel looked after. But persuasion without definition can collapse under compression.
Before rewriting, ask what pressure the answer surface applied. Did it demand a definition? A comparison? A local signal? A process? A credential? A constraint? Then inspect whether the page answered that pressure or slid away into generalities.
This is not only a search exercise. A buyer applies similar pressure, though less visibly. The HR manager comparing training providers wants to know which programme fits their team and compliance burden. The public-sector department wants to know whether procurement and delivery constraints are understood. The technology firm wants to know whether a trainer can handle a mixed-experience cohort. If the page says “flexible learning” to all three, it may sound welcoming while telling none of them enough.
The best rewrite after an AI Overview audit often feels like a correction of manners. The page stops talking around the visitor and starts answering the question in the room.
The Rain Check — Window: an AI Overview that chose training pages with plainer structure over a smoother service brand. Grain: the repeated winner was delivery detail placed near audience fit and trainer proof. Umbrella: read the answer shape first, then rewrite only the claim, proof, or section that failed compression. Last Drop: Rain shows the dip in the pavement before anyone admits the road was uneven.