SEO content writing when the answer engine gets there first

What actually changed

The story usually told is that AI summaries arrived and organic traffic collapsed. The truth is less dramatic and more useful to understand, because the direction of travel started well before generative summaries appeared.

Search results have been getting progressively better at answering simple questions without a click for over a decade. Featured snippets, knowledge panels, calculators, currency conversion, weather, sports scores, definitions — every one of those removed a category of query from the pool that ever produced a visit. Generative summaries accelerated a trend rather than starting one.

What this means practically is that the queries which still send traffic are the ones where a summary is not a satisfying answer. That is a shorter list than it used to be, but it is not a short list:

• Queries where the searcher needs to evaluate and choose, not merely know
• Queries where trust in the source is part of the answer — health, money, legal, safety
• Queries requiring proprietary data that no model was trained on and no summary can synthesise
• Queries where the searcher wants a tool, a template, or a file, not prose
• Queries with genuine local, temporal or personal specificity
• Queries where opinion and judgement are the value — “is X worth it for a team like mine”

Notice how many of those overlap with the commercially valuable end of the funnel. The queries losing clicks fastest are disproportionately the ones that were never worth much anyway: definitional, top-of-funnel, high-volume, low-intent. Many content programmes were built almost entirely on those, which is why the pain has been unevenly distributed.

The right question is no longer “how do we rank for this?” It is “if a machine summarises this perfectly, is there any reason left to visit us?”The filter we apply to every brief

Intent is still the whole game

Every durable improvement in search performance we have made for a client traces back to matching intent more precisely than the competition, not to any technical trick.

The practical method is unglamorous. Take the query. Open the results. Do not look at the rankings — look at the shape of what is ranking. Are they listicles? Product pages? Forum threads? Videos? Tools? That shape is the clearest available statement of what searchers accepted as a useful answer.

Then ask the harder question: what is the job behind the query? “Best project management software” is not a request for a list of project management software. It is usually a request for help with a decision the searcher does not feel qualified to make. A list satisfies the query. A decision framework satisfies the job — and is far more likely to be remembered, cited, and returned to.

The four intent mismatches we see most

1. Answering a comparison query with a sales page. Somebody comparing you to a competitor wants an even-handed comparison. Give them a genuinely fair one, including where the competitor wins. It converts better than a one-sided page, and it is the only version that survives scrutiny.
2. Answering a how-to query with a why-it-matters essay. Somebody wants steps. Give steps first; context can come after.
3. Answering a definitional query with 3,000 words. If the searcher wants a definition, lead with a two-sentence definition. Everything else is optional depth for the minority who want it.
4. Answering a tool-shaped query with prose. If people want a calculator, build the calculator. Prose about the calculation will not compete.

Structure your page for extraction

Whether a summary engine, a snippet algorithm or a human skimmer is reading, the same structural habits help. Content that is easy to extract from is content that gets used — and citation in a summary is now a meaningful acquisition channel in its own right.

• Answer in the first hundred words. The direct answer, stated plainly, before context. Then earn the reader's attention for the depth beneath it.
• One idea per heading, phrased as the question it answers. Headings are the table of contents an extraction system reads.
• Self-contained paragraphs. A paragraph that only makes sense after reading the previous three cannot be quoted. Assume any paragraph might be lifted alone.
• Put facts in tables and lists. Structured data is disproportionately likely to be extracted and cited.
• Be specific with numbers and dates. “Recently” is unusable. “In March 2026” is quotable.
• Attribute clearly. Where a claim comes from your own data, say so explicitly in the sentence — that attribution is what carries your brand into a summary.

Depth beats length, and they are not the same

The correlation between word count and rankings was always a correlation with something else: comprehensive pages tend to be longer, so longer pages tended to rank. Teams responded by padding. Padding does not work, and it never really did.

Depth means covering the questions a reader will have next. The reliable way to find them is to write the piece, then list every question a sceptical reader would ask after each section. Answer the ones that matter. That process adds words, but it adds them for a reason.

Signals that a page has length without depth:

• A history section nobody asked for
• Restating the question in three different ways before answering
• Definitions of terms your audience already knows
• An introduction explaining why the topic is important
• A conclusion that summarises what was just read

Cut all five and most pages lose a quarter of their words and none of their value.

Original input is the durable moat

If a page can be assembled entirely from other pages, a model can assemble it faster, and there is no reason for it to send anyone to you. The defensible material is whatever cannot be synthesised from existing text.

In rough order of accessibility:

1. Your own product or operational data, aggregated and anonymised. Usually the cheapest original asset a company owns and almost always underused.
2. Original testing. Buy the things, run them, publish the results with methodology. Expensive, extremely durable.
3. Practitioner experience. What actually happens when you do this at scale, including the failures. Nearly impossible to fake convincingly.
4. Primary interviews. Named people saying specific things on the record.
5. Surveys of your own audience. Cheap, quotable, and endlessly refreshable year on year.

A programme that publishes two genuinely original pieces a quarter and supports them with well-executed conventional content will out-perform one publishing twelve conventional pieces a month. We have watched this play out often enough to state it plainly.

The technical floor you cannot skip

Technical work rarely wins on its own, but it reliably loses. The floor:

• Crawlability and indexation. Check what is actually indexed rather than assuming. Orphaned and accidentally excluded pages are more common than anyone expects.
• Sensible internal linking. Cornerstone pieces should be linked from every supporting piece using descriptive anchors. This is the highest-leverage on-site work available and costs nothing.
• Fast, stable pages. Particularly layout stability — content that jumps as it loads is punished by readers before any algorithm gets involved.
• Structured data where it is genuinely applicable. Article, FAQ, HowTo, Organisation, BreadcrumbList. Do not mark up things that are not there.
• Clean, stable URLs. Changing them without redirects remains one of the most reliable ways to destroy a year of work in an afternoon.
• One canonical version of every page. Duplicate paths quietly split every signal you earn.

Updating: the highest-return work in search

Across our client base, refreshing existing content consistently returns more per hour invested than producing new content — usually by a wide margin. It is also the least popular work on any team, because nobody gets credit for it.

A workable quarterly routine:

1. Pull every page with a meaningful decline over the last six months.
2. Sort by how much traffic they used to get, descending.
3. Take the top ten and read them properly. Not skim — read.
4. Classify each: stale facts, superseded by better competitors, intent has shifted, or thin from the start.
5. Fix the first three categories. Consolidate or delete the fourth.

That last step matters more than it looks. Deleting genuinely weak pages, with redirects to their nearest useful equivalent, has produced site-wide improvements for several of our clients. A site is judged in aggregate; carrying two hundred pages nobody should read is a cost, not an asset.

Measuring when clicks are no longer the point

If summaries answer more queries, impressions can rise while clicks fall, and a click-only dashboard will report failure during a period of growing influence. Broaden what you watch:

The open-text “how did you hear about us” field deserves particular attention. It is the least sophisticated measurement instrument available and frequently the most accurate one, because it asks the person who knows.

The pre-publish checklist

Eleven items. If a piece fails more than two, it is not ready.

• The primary question is answered in the first hundred words
• The format matches what is already satisfying that query
• At least one element is genuinely original to us
• Every heading is a question a reader would actually ask
• Every paragraph makes sense read alone
• Facts and comparisons are in tables, not buried in prose
• Claims from our own data say so inside the sentence
• Numbers and dates are specific
• Internal links point to the relevant cornerstone with descriptive anchors
• Nothing in it exists purely to increase word count
• A subject expert has read it and did not wince

That last one has no measurement attached and is the single best predictor we have of whether a piece will perform.