Trend articles are cheap to write because nobody has to act on them. You can assert that AI will change everything, publish it, and never be measured against it.

So this is structured differently. Each shift below is paired with a specific change we have made to how we work, what it cost, and where we are still unsure. If a trend did not produce a change, it did not make the list.

1. Answer engines are absorbing informational search

What we see: across our own site and client sites, informational queries increasingly resolve without a click. Definitional content — "what is X", "how does Y work" — draws less traffic than equivalent content did two years ago, while the pages that hold up are the specific ones: comparisons, costed breakdowns, decision frameworks, anything with a number or a named trade-off in it.

What we changed:

• Every article now opens with a direct, self-contained answer of roughly forty to sixty words, written so it still makes sense when quoted alone. The blog platform has a dedicated field for it rather than leaving it to the writer's discretion.
• We stopped commissioning definitional content unless it is a genuine prerequisite for something specific that follows it.
• We added structured data — article, breadcrumb and FAQ schema — as a build requirement rather than an optimisation task.
• We publish the numbers. The cost breakdowns in what a CRM implementation actually costs exist partly because a specific figure with stated assumptions is something an aggregator cannot manufacture.

What it cost: roughly a 30 percent increase in production time per article, and a reduction in publishing volume. We publish fewer, longer, more specific pieces than we did.

Still unsure: whether citation in an answer engine converts at any useful rate. We can see that it happens. We cannot yet reliably connect it to pipeline, and anyone claiming otherwise with precision is guessing.

2. Buyers self-serve much further before contact

What we see: first conversations start later and better informed. Prospects arrive having read pricing pages, comparison content and case studies, often ours and three competitors'. The discovery call that used to establish basics now starts at objections.

What we changed:

• We publish cost models rather than "contact us for pricing". Not fixed prices — the honest version is a model with stated assumptions, which is more useful and more defensible.
• We write for the objection, not around it. The case against redesigning your website exists because the buyer is already having that argument internally.
• We shortened discovery calls and moved the qualification questions into the form, so the conversation starts where the buyer already is.

What it cost: fewer enquiries, higher conversion from enquiry to proposal. Publishing cost models filters out buyers looking for the cheapest possible number, which is uncomfortable in a slow month and correct over a year.

Still unsure: where the line sits between useful transparency and giving away the analysis that clients pay for. We have moved it further toward transparency twice and have not yet regretted it.

3. Tracking decay is pushing measurement back to first-party data

What we see: client-side measurement keeps getting less complete — browser restrictions, blockers, consent requirements. The gap between what happened and what your analytics recorded widens quietly, and it is not evenly distributed across audiences.

What we changed:

• Server-side event collection is now the default in our proposals for anything commercially meaningful, not an upsell.
• Every engagement that touches measurement starts with a tracking plan document, in version control, before any implementation.
• We report identity match rate to clients explicitly, including when it is bad. It is usually worse than anyone expects and the number changes decisions.
• We push modelled conversions — qualified opportunities, revenue — back to ad platforms instead of form fills wherever the data foundation supports it.

What it cost: proposals got more expensive and harder to sell. The foundation work in the data plumbing nobody budgets for adds real money to a campaign engagement before a single ad runs. We lost work over it. We still think it is the right call, because the alternative is optimising against a signal we know to be wrong.

4. AI compresses building, not deciding

What we see: in our own delivery, AI assistance has genuinely reduced implementation hours — writing code, first-draft content, configuration, test scaffolding. It has barely touched discovery, architectural decisions, review, integration debugging or stakeholder alignment.

What we changed:

• Our estimates changed shape rather than shrinking. Build lines came down; discovery, review and testing lines went up in proportion and slightly in absolute terms.
• Review became a formal, budgeted stage. More code produced faster requires more review, not less, and pretending otherwise is how defect rates climb.
• We stopped charging by output volume for content. When drafting is cheap, paying for drafts is paying for the wrong thing.
• We now say plainly in proposals which parts of the work are AI-assisted and which are not, because clients ask and evasiveness reads badly.

What it cost: retraining time, and a genuinely awkward period of re-estimating projects while our own historical benchmarks were no longer valid. We over-promised speed on two projects during that period and had to absorb the difference.

Still unsure: whether the review burden keeps scaling with generation speed, or whether tooling closes that gap. Our current planning assumes it does not.

5. Tool consolidation is finally happening

What we see: after years of accumulation, buyers are actively cutting. Consolidation requests now arrive as often as new-tool implementation requests, which was not true two years ago.

What we changed:

• We built stack audits into a fixed-scope service rather than treating them as pre-sales work.
• We now ask what a client will switch off before recommending anything new. If the answer is nothing, we usually recommend nothing.
• Our own stack went from 31 tools to 19 last year. Doing it to ourselves first was informative and mildly humbling — we found three tools nobody could account for.

What it cost: some implementation revenue, replaced by consolidation revenue at lower ticket sizes and higher trust. The relevant argument is in most companies own 40+ martech tools and use 12.

6. Undifferentiated content has stopped working

What we see: generic content is now effectively free to produce, which means its value has converged on its cost. What holds up is content that requires something an aggregator cannot obtain: original data, real project numbers, a position with a cost attached, or an account of a failure.

What we changed:

• We publish roughly a third of the volume we did and put the difference into depth.
• Every piece must contain at least one of: a number we can defend, a position we would lose work for, or a process nobody else has documented.
• We started publishing post-mortems on work that underperformed. It is the least comfortable change on this list and the one that has produced the best conversations.
• Author attribution and review dates are now required fields on every article, because who wrote something is part of whether it should be trusted.

What it cost: volume-based organic growth slowed for about two quarters before the depth started compounding. If your reporting is set up to reward publishing frequency, this change will look like failure for longer than is comfortable.

Three things we might be wrong about

We may be over-rotating on answer engines. The share of commercially valuable queries they absorb is still small compared to the share of informational ones. If high-intent search behaviour stays stable, the traditional playbook survives longer than we are assuming.

We may be underestimating how fast AI closes the judgement gap. Our estimating model assumes discovery and review stay human-intensive. That assumption is roughly a year old and we intend to re-test it rather than defend it.

Transparency may have a limit we have not found yet. Publishing cost models and post-mortems has been net positive so far. It is possible that continuing produces a point where prospects use the material and never make contact. We will know before we can prove it, and we will say so here when we do.

If you are making similar changes, or think one of these is wrong, we would genuinely like to hear it — most of this list came out of conversations rather than analysis. Get in touch, or see how the thinking shows up in what we actually do.