The short version: Most AI news is irrelevant to your growth numbers. In this launch edition of AI Signal, we cut the model-release noise down to five developments that actually change how growth teams operate this quarter — cheap long-context models that make research and personalization economical, agentic browsing that breaks parts of your funnel analytics, AI-generated content getting commoditized (and what replaces it as a moat), the collapse of build-vs-buy timelines for internal growth tooling, and the arrival of hard data showing most enterprise AI pilots don't pay back. Below each item: what changed, why it matters to a growth team, and what we'd actually do about it in the next 30 days.
AI Signal is Growaton's biweekly roundup. We're not here to tell you a new model scored three points higher on a benchmark. We're here to answer one question every time: does this change what a growth team should ship next week? If the answer is no, it doesn't make the list.
Why we're launching an AI news roundup for growth teams
There is no shortage of AI news. There is an enormous shortage of AI news filtered through the lens of "will this move CAC, conversion rate, activation, or payback period?"
Most AI coverage is written for one of three audiences: investors tracking model capability curves, engineers evaluating architecture decisions, or generalist readers who want to know if their job is safe. Growth teams — the people who own the number — get almost nothing. So they end up either ignoring AI entirely, or chasing every shiny release and burning a quarter on tooling that never touched revenue.
We run growth pods for seed-through-Series-C SaaS, fintech, marketplace, and e-commerce companies. We ship weekly. That means we're stress-testing AI tooling in production against real conversion metrics constantly, and we see the gap between what AI demos well and what actually survives a month of real traffic. That gap is what this stream is about.
Our filter for every item in AI Signal:
| Filter question | Why it matters |
|---|---|
| Does this change unit economics? | If it doesn't move cost-per-output or output-per-hour, it's a curiosity. |
| Can a 5-person growth team use it in under 2 weeks? | Enterprise-only capability is not actionable for a Series A. |
| Does it break something we currently rely on? | Analytics, attribution, and SEO all have AI-shaped cracks forming. |
| Is there evidence, or just a demo? | Vendor demos are marketing. Production data is signal. |
Signal 1: Long-context models got cheap enough to change your research workflow
What changed
The price-per-token curve for frontier-adjacent models has fallen dramatically while context windows expanded into the hundreds of thousands to millions of tokens. Google's Gemini documentation lists context windows in the 1M-token range for its long-context models, and OpenAI's pricing page shows a spread of several orders of magnitude between its cheapest small models and its frontier reasoning models.
The practical consequence: tasks that were economically absurd 18 months ago — "read all 340 of our support tickets from last quarter and cluster the objections" — now cost less than lunch.
Why it matters for growth
This is the single most underexploited AI capability in growth right now, and it has nothing to do with content generation.
The bottleneck in most growth programs isn't execution speed. It's hypothesis quality. Teams run mediocre experiments because nobody has time to read 400 sales call transcripts, 2,000 churn survey responses, 900 onboarding session recordings, and six months of Intercom conversations to find the actual friction. So they test button colors instead.
Cheap long-context inference removes that constraint. We've used this pattern in pods to turn qualitative goldmines into ranked experiment backlogs:
- Dump all closed-lost CRM notes from two quarters into a long-context model, ask it to cluster loss reasons and quantify each cluster, then cross-reference against your pricing page copy.
- Feed 100+ onboarding support tickets and ask which step in the flow generates the most confusion per user reaching it — not in aggregate, but normalized by traffic.
- Concatenate every competitor's pricing page, docs site, and G2 review corpus and ask for the objections your messaging doesn't currently address.
The output isn't an answer. It's a prioritized list of things worth testing — which is exactly what the Diagnostics phase of a growth program is supposed to produce, and what usually takes three weeks of analyst time.
What we'd do in the next 30 days
Pick your single richest unstructured data source (usually sales call transcripts or support tickets). Run one long-context synthesis pass. Convert the top three findings into experiment hypotheses with a stated success metric. Total cost: a few dollars in tokens and about four hours of a senior person's attention.
Signal 2: Agentic browsing is quietly breaking your funnel data
What changed
AI agents that browse, click, fill forms, and complete tasks on behalf of users moved from research demos to shipped products across major vendors in 2025. Meanwhile, referral traffic from AI assistants has become a measurable channel — and one that behaves nothing like organic search.
Why it matters for growth
Three concrete problems, all of which will show up in your dashboards before anyone tells you why:
1. Zero-click research is compressing top-of-funnel traffic without compressing demand. Users get their comparison answer inside an AI assistant and only visit your site when they're much closer to a decision. Net effect: sessions down, conversion rate up, and every historical benchmark you had for "healthy traffic" is now wrong. If you're grading your SEO program on sessions, you'll fire a working program.
2. Agent traffic pollutes behavioral analytics. Agents don't scroll like humans, don't hesitate like humans, and sometimes fill forms with placeholder data. Bot filtering built for 2019-era scrapers doesn't reliably catch them. Your bounce rate, time-on-page, and heatmap data get noisier, and any A/B test with low absolute volume gets noisier faster.
3. Your funnel may need to be machine-legible. If a meaningful share of buyers are researching through an assistant, then whether your pricing, integration list, and security posture are stated in plain, crawlable, unambiguous text matters as much as your hero animation. Content locked behind JS rendering, tucked in PDFs, or implied rather than stated simply doesn't exist to a retrieval system.
What we'd do in the next 30 days
Three things, in order:
- Segment AI referral traffic explicitly. Create a channel grouping for known assistant referrers and look at its conversion rate separately. In our client work this segment often converts at multiples of generic organic — which changes how you value the content that earned it.
- Audit for machine legibility. Are your prices stated as numbers in HTML? Is your integration list a crawlable page or a logo carousel? Is your security/compliance status written out or a badge image?
- Re-baseline your traffic KPIs. Shift primary SEO reporting from sessions to qualified pipeline per topic cluster. If your reporting layer can't do that, that's a data infrastructure problem worth fixing before it costs you a good channel.
Signal 3: AI content generation is now table stakes — the moat moved to distribution and proof
What changed
Generating competent prose is solved and effectively free. Every competitor in your category has the same capability at the same price. Simultaneously, Google's guidance has stayed consistent: it rewards helpful, reliable, people-first content regardless of how it was produced, and its spam policies explicitly target scaled content abuse — content produced primarily to manipulate rankings rather than help people.
Why it matters for growth
The obvious inference — "AI content doesn't work" — is wrong. The correct inference is that AI-generated content has no defensible differentiation on its own, because the marginal cost of production went to zero for everyone at once. When supply of adequate content is infinite, the scarce inputs become:
- Proprietary data. Numbers only you have: benchmarks from your product, aggregate customer behavior, experiment results.
- First-hand experience. What actually happened when you tried the thing, including where it failed.
- A point of view. An actual opinion that someone could disagree with.
- Distribution. Whether anyone with authority will link to, cite, or share it.
Here's the framing we use with clients:
| Input | Cost trend | Defensibility |
|---|---|---|
| Competent writing | → $0 | None |
| Keyword research & topic coverage | → near $0 | Low |
| Proprietary benchmark data | Flat / expensive | High |
| Documented experiment results | Flat / expensive | High |
| Practitioner point of view | Flat | High |
| Earned links & citations | Rising | High |
AI's honest job in a content program is removing the drafting and formatting tax so senior people spend their hours on the high-defensibility inputs. Used that way, it's a genuine multiplier. Used to publish 200 undifferentiated posts a month, it's a liability that also happens to be against the guidelines.
What we'd do in the next 30 days
Audit your last 20 published pieces against one question: does this contain at least one fact, number, or opinion that no competitor could publish? Whatever fails that test isn't worth updating — it's worth consolidating or removing. Then instrument your product or CRM to produce one recurring proprietary statistic you can publish quarterly. That single asset will out-earn a year of generic posts.
Signal 4: The build-vs-buy math flipped for internal growth tooling
What changed
AI-assisted development has substantially compressed the time to build small, purpose-specific internal tools. This is now measurable rather than anecdotal: a randomized controlled trial by GitHub found developers using Copilot completed a specific coding task 55% faster than the control group, and McKinsey's research on generative AI in software engineering reported large speedups on code generation and refactoring tasks, with much smaller gains on high-complexity, unfamiliar work.
Note the shape of that finding: big gains on well-specified, bounded, familiar problems; small gains on genuinely hard novel work. Internal growth tooling is almost entirely the former.
Why it matters for growth
Growth teams accumulate SaaS subscriptions because building was slow. A lead-routing rule that your CRM can't express, a churn-risk scorer, a custom onboarding checklist, an internal dashboard that joins product events to billing data — each was a two-month engineering ticket nobody would approve, so you bought a $1,200/month tool that does 60% of what you need.
When the same thing takes a senior engineer four days, the calculus changes. And critically, the custom version does 100% of what you need and owns its own data.
Where we've seen this land hardest in pods:
- Internal experiment infrastructure — assignment, exposure logging, and results readout wired to your real revenue data instead of a vendor's event schema.
- RevOps glue — the enrichment, deduping, and routing logic that sits between four systems and currently lives in a Zapier account nobody understands.
- Purpose-built onboarding surfaces — in-product checklists and activation nudges instead of a generic overlay tool.
- Reverse-ETL and warehouse-native activation — pushing product signals into your ad platforms and CRM without a per-event pricing tier.
This is the practical core of the "senior builders at AI speed" thesis and the reason our pods carry engineering rather than partnering with an agency for it. When growth engineering and marketing sit in the same standup, "we should test that" becomes shipped code in the same week instead of a Q3 roadmap item. If you want the mechanics, we've written them up in our growth engineering approach.
What we'd do in the next 30 days
List every growth-adjacent SaaS tool costing over $500/month. For each, write one sentence describing what it actually does for you. Any tool whose one sentence is simple enough to build is now a build-vs-buy conversation, not a renewal.
Signal 5: The ROI evidence arrived, and most AI pilots are failing it
What changed
The uncomfortable data point of the past year: research from MIT's NANDA initiative, widely reported in mid-2025, found that roughly 95% of enterprise generative AI pilots produced no measurable P&L impact. Separately, S&P Global Market Intelligence reported a rising share of companies abandoning AI initiatives before production.
Meanwhile, aggregate adoption keeps climbing — McKinsey's global survey shows a large majority of organizations now using gen AI in at least one function, but far fewer reporting meaningful EBIT impact at the enterprise level.
Why it matters for growth
Read those together and the story is clear: adoption is not the hard part; attribution and integration are. The pilots that fail tend to share a profile:
- Started from the tool, not from a bottleneck ("we should use AI for something").
- No baseline captured before rollout, so improvement is unprovable.
- Success measured in activity (prompts sent, drafts produced) rather than outcomes.
- Never integrated with the systems where the work actually happens, so humans do double entry.
- Owned by nobody after the demo.
The pilots that work look boring by comparison. They pick a single expensive, high-volume, well-specified workflow. They measure the before-state in hours and dollars. They wire the AI step into existing systems so there's no parallel process. They assign one owner. They report payback in the same terms as any other investment.
We treat AI spend exactly like paid media spend: it needs a baseline, an attributable output, and a payback period. If you can't state what an AI workflow costs per month and what it produces in hours saved or revenue influenced, you don't have an AI program — you have a subscription. That discipline is the whole reason we built out our work on proving LLM ROI, and it's the first thing we look for in a diagnostic.
What we'd do in the next 30 days
Take your existing AI spend — every seat, every API key, every tool — and put it in one table with three columns: monthly cost, the specific workflow it serves, and the measurable output. Anything with a blank third column gets a 30-day deadline to produce one or gets cancelled. Most teams find 30–50% of their AI line item is unattributed.
The through-line: AI changes velocity, not judgment
Five signals, one pattern. Every development above increases how fast you can do things. None of them tells you which things to do.
That's the failure mode we see most often. Teams adopt AI, triple their output volume, and see no change in their numbers — because they tripled the output of a strategy that wasn't working. Faster execution against a bad hypothesis just gets you to the wrong answer sooner and with a bigger invoice.
The teams compounding real advantage from AI right now share three traits:
- They start from bottlenecks, not tools. The question is "where is our growth constrained?" and only then "can AI relieve that constraint cheaper than a hire?"
- They keep judgment senior and human. AI drafts, clusters, summarizes, and builds. Humans decide what to test, what the result means, and what to do next. That ratio doesn't invert as models improve — it gets more valuable, because everyone else's execution speed is rising too.
- They measure ruthlessly and kill fast. Baseline, instrument, review on a fixed cadence, cancel what doesn't pay back. Same rigor as any experiment-driven growth program.
Nothing about that is new. AI just raised the penalty for skipping it.
What we're watching for the next edition
Candidates already on the board for AI Signal #2:
- Agent-mediated purchasing. As assistants move from research to transacting, checkout and pricing pages become interfaces for machines as well as humans. Early days, big implications for e-commerce and self-serve SaaS.
- Real-world numbers on AI SDR and outbound tooling. Lots of vendor claims, very little independent reply-rate and pipeline data. We'll publish ours.
- Inference cost trajectory vs. reasoning-model cost. Cheap models keep getting cheaper while reasoning models are expensive per call. Routing between them intelligently is becoming a real unit-economics lever.
- Evaluation as a growth competency. Teams shipping AI features in-product need eval harnesses. Almost none have them. This is going to be a visible quality gap in 2026.
If there's something breaking or working in your stack that you think belongs in the next edition, tell us. Practitioner reports beat press releases every time.
How to act on this without hiring an AI team
You don't need an AI strategy. You need a growth strategy where AI is one of the levers, evaluated on the same terms as everything else.
Practically, that means three capabilities in the same room: someone who can read the data and find the real bottleneck, someone who can build the thing, and someone who owns the distribution and conversion side. When those three sit in different companies — or different quarters of a roadmap — AI speed doesn't help you, because handoff latency dominates execution time.
That's the model we run: one embedded pod covering product, engineering, data, experimentation, and marketing, shipping weekly against a named metric. If you want to see what that looks like against real numbers, our case studies walk through specific experiments — including the ones that lost — and our pod engagement options lay out how scope works.
And if you'd rather just get an outside read on where your growth is actually constrained and whether AI is anywhere near the answer, book a free growth diagnostic. It's a conversation, not a pitch deck. Worst case you leave with a prioritized list of bottlenecks.
