Convert.com

The WiE Summit 2026 line-up drops tomorrow. Many of these names haven't been on a big stage before. That's the point. 121 people applied. Ten got a slot. The three-judge panel wasn't looking for who's already famous, but for who has something the field hasn't heard yet. Some of these speakers are veterans. Others are spreading their wings. What all of them share is a talk that couldn't have been given at any other event this year. It bets on people doing the sharpest work, coaches them, pays them, and puts them on a stage, all to make the CRO, growth, and UX space a little richer. Tomorrow, we're introducing them to you. 9-10 September 2026, online, free to attend, and 72-hour replays for everyone who registers.

100% of senior CROs use AI. Only 9% get an edge. For the rest, it's hallucinations, broken workflows, and token bills. Tomorrow's workshop is for them. Iqbal Ali points to the culprit: the non-determinism of generative AI. Ask a model the same question multiple times and you get different answers. He's covering four things tomorrow: 1. Small language models and how to pick the right one 2. Building serialized workflows on n8n and defining their sequence 3. Validating claims and code quality deterministically 4. Where small models fall short (and what to do instead) Slides are scarce. The session covers the actual techniques and principles behind the workflows. Tomorrow at 10am EDT, hosted with Ressada and Experimentation Elite. Replay goes to everyone who registers. Grab a seat: us06web.zoom.us/webinar/regist…

Frontier models won't fix a broken AI workflow. Small models don't need to. Teams are still paying frontier prices to ship demos, not products. They get non-deterministic output and a token bill that scales linearly with usage. Iqbal Ali built Ressada on the other approach: small open models, serialised workflows, deterministic checks between every step. The stack is cheaper to run than a single frontier API call, and it doesn't drift. The technique comes from software engineering discipline that predates AI by decades. Frontier-model marketing has convinced most teams to forget it. Iqbal's workshop on July 28 is where he covers this in detail, hosted with Ressada and Experimentation Elite. Free to attend. He starts with the principles, then breaks down his actual Ressada stack: how each piece is put together, what runs where, and why the bill stays under control. Come along if you're tired of paying frontier prices for output you can't trust. us06web.zoom.us/webinar/regist…

Your AI workflow is bleeding tokens and delivering different answers every run. Both problems have the same fix. We're running a workshop on July 28 where Iqbal Ali walks through it, with Ressada and Experimentation Elite. The dependency is what makes both problems worse. When your workflow relies on one big-tech model, you inherit its non-determinism, its price hikes, its API changes, and its hallucinations at the same time. The alternative: 1. Break the workflow into stages 2. Apply rule-based logic where it fits 3. Bring in a small model where it doesn't 4. Validate deterministically between each step These techniques come from software engineering and predate AI by decades. Iqbal is showing how to apply them to modern AI outputs on n8n. Result: reliable output at a fraction of the cost, and nothing depends on someone else's black box. Save your seat: us06web.zoom.us/webinar/regist…

Iqbal Ali is teaching the framework for reliable AI workflows on July 28. By the end, you'll know how to build six specific pieces of a working stack. The core argument is that reliable AI workflows come from architecture. That means small models on n8n, deterministic validation between steps, and each stage returning the same output on every run. By the end of the webinar, you'll know how to build: - Reliable n8n workflow with deterministic output - Framework for choosing the right small language model - A/B test code validated deterministically - Claim checker for convincing-but-empty content - Regex proven correct by tools - Much smaller token bill for enterprise-grade output July 28 at 10am EDT, hosted with Ressada and Experimentation Elite. If your AI workflows are hallucinating, breaking, or costing hours to check, block off the hour. Link to register in the comments.

Ask an AI 2+2 a hundred times, and you get "4" in ninety-nine different ways. Which is why your workflow breaks. Iqbal Ali just explained why AI workflows keep breaking in the clearest 30 seconds we've heard. LLMs are non-deterministic. Ask them the same question multiple times, and the answer stays right while the format changes. You might get "four," "the answer is 4," "IV," or a full sentence like "two plus two equals four." Your automation pipeline treats each of those as a different output. And the variability compounds every time the prompt gets more complex. The failure point is format. Output shape changes on every call, and anything downstream that expects a consistent shape breaks with it. Iqbal is teaching the fix on July 28, hosted with Ressada and Experimentation Elite. Small language models, n8n workflows, and the validation patterns that make each step deterministic. If your AI workflows are hallucinating, breaking, or costing hours to check, join us on July 28 at 10am EDT. us06web.zoom.us/webinar/regist…

MCP hype is ignoring one detail. The AI on the other side of those tools is still non-deterministic. Ask an AI model the same question twice, and you get two different answers. That is a manageable problem when the AI is generating text. It becomes a very different problem when the AI has direct MCP access to real tools. Giving an AI MCP access is handing it the tools to cause havoc, then trusting it won't. And the AI you handed them to will not behave the same way twice. Iqbal Ali's fix has two parts. MCPO. It converts MCPs back into a well-documented API service with clean, structured endpoints. You get the ease of MCPs and the specificity of APIs, without blind trust in whatever the model decides to do. Then small language models running through serialized workflows on n8n. Each step is deterministic. You can validate the output at every stage. Iqbal's workshop on this is July 28, hosted with Ressada and Experimentation Elite. If your AI workflows are hallucinating, breaking, or costing hours to check, join us at 10am EDT. Worth attending live if you can: us06web.zoom.us/webinar/regist…

AI adoption in CRO hit 100% 9% get value Using AI is not enough. Convert's in-house research shows that 100% of seasoned CROs (in agency settings) use AI tools. Yet only 9% believe their implementation lends them an edge. For the rest? AI intensity comes with hallucinations, downtimes, workflows that break, and outputs you spend twice as long checking for corruption and mistakes. Veteran AI specialist @iqbalacomics points to a culprit: The non-determinism of generative AI! [Ask an AI model the same question multiple times and you will receive different answers] In this session, Iqbal will share the foundational knowledge and tools you need to autonomously implement a grounded AI infrastructure while drastically reducing token costs. We are covering: 1. Small language models and how to pick the right one 2. Benefits of building serialized workflows on platforms like n8n (and defining their sequence) 3. Validating claims and code quality deterministically 4. Where small models fall short (and what to do instead) Slides are scarce! The session is designed to talk about the techniques and principles used in the workflows. Limited-time replay will be made available. **PS: Want to build the next big thing with AI, the right way? We're kicking off a series of events for a new generation of builders to learn from previous generations' experts! AI works, when you know what you're doing. us06web.zoom.us/webinar/regist…

GA4 can take 20 seconds to send its first pageview on slow 3G. Mobile shoppers leave in three. We spent nearly two years debugging GA4 vs A/B test data mismatches across dozens of live customer sites. The finding: most gaps above 5-10% are not configuration issues. They come from your slow-device, slow-network visitors. Three patterns showed up over and over: 1. The 20-second first-pageview delay. Under slow 3G or 4G conditions, GA4's first page_view frequently fires only after 20 seconds or more. That's enough for a mobile shopper to open your page, skim it, and leave. GA4 never registers the session. Your A/B test tool, running synchronously in the head, already logged the visit. This isn't documented anywhere by Google. 2. A funnel shape that gives it away. If the loss were random, session and conversion counts would drift by the same percentage. They don't. The biggest GA4 vs A/B test gaps show up at funnel entry and final conversion. Mid-funnel looks almost fine. Every GA4 event ties to a session, and every session ties to a page_view. Miss the first page_view, and every downstream event from that visitor is missing too. Visitors who got past the first 20 seconds already cleared the loading bottleneck, so their mid-funnel events fire on time. 3. The GTM "Window Loaded" trap. Common advice to fire GA4 tags on Window Loaded instead of Page View improves Core Web Vitals. It also makes analytics worse. Page speed reports look better. Slow-network sessions get missed. Only about half of US connections are on 5G as of 2025, and "on 5G" often means 4G or worse in practice. Slow-device visitors are not a rounding error. The next time GA4 disagrees with your A/B test tool by more than 5-10%, pull the funnel from both. If the gaps are largest at top and bottom and small in the middle, you are seeing the slow-network signature. If the gap is even across all stages, it is likely configuration or a real behavioral difference. Full 7-step diagnostic in the article. convert.com/blog/a-b-testi…

TikTok just voted against AI content. By banning it from the only place that has to sell. TikTok Shop's new quality rules ban AI voices, still images, static screens, and PDP screenshots from lives and shoppable content. The rules prioritise: - Real-time human presenters - Multi-angle demos - Dynamic footage Everything TikTok banned in the shop is still allowed (and often rewarded) on the main feed. AI-generated voiceovers, static image slideshows, and screenshot-based content thrive on the FYP. The delineation is worth noticing. When TikTok's main product is engagement, AI content is welcome. For actual sales, humans win. One question everyone in commerce content should be asking: will these rules eventually reach the FYP?

Iqbal Ali built an AI system that writes, tests, and self-heals A/B test JavaScript for under $0.06 per run. Or free if you run it locally with Ollama. The workflow takes two inputs: a page URL and a plain-English description of the change you want to test. From there, it does eight things in sequence: 1. Fetches the HTML of the target page via Browserless. 2. Generates a technical specification (what a senior dev would give a junior dev). 3. Identifies the QA rules needed to validate the change. 4. Writes the JavaScript. 5. Generates JavaScript tests to check the change. 6. Sends everything to Browserless to render the change and run the tests. 7. Runs a self-healing loop (up to 3 attempts) to fix any issues the tests catch. 8. Returns the specification, QA rules, JavaScript, and test results, ready to plug into your testing tool. Three techniques from the article worth pulling out for anyone building AI workflows: Think about validation before generation. The specification and QA rules get written first, before any code, so the output has a way to be validated. Use deterministic tests. The QA tests are written in JavaScript, not run by an AI agent. Deterministic code checking non-deterministic output is what makes the workflow trustworthy. Serialised workflow beats agent chaos. Every step has one job. The rigid structure is what makes the whole thing reliable enough to trust. This is a friction reducer, not a developer replacement. When code generation costs six cents, more test ideas actually get built. youtu.be/BANVRGp8aZo?si…

Andrea Bronzini ran the same A/B test 100 times. What happened to 88 of them should change how you read every result. Andrea is the founder of Confident Story. The way he thinks about noise in A/B tests is where the interview gets interesting. His 5-word definition of optimisation: "A balancing act between errors." The interview lands on three ideas worth pulling out: 1. There are three failure modes in A/B testing, not one. The industry has been optimising for one: calling the wrong winner (false positives). Statistical significance handles that. The second failure: missing real winners because their lift didn't cross 95%. "Most winners get lost in the inconclusive pile, and nobody counts that as a cost." And the third failure: shipping inflated winners. "You called the right direction, but your measured lift was 12% and the true lift was 4%." The three are structurally intertwined. Make one threshold stricter and the other two get worse. 2. The meta-experiment that proves it. Andrea built a simulator that replays the same A/B test 100 times with a known true lift of 3.2%. Only 12 out of 100 crossed 95% significance. Those 12 measured lifts between 9% and 14%. "Winners don't disappear after launch because of implementation problems or seasonality. They disappear because you never shipped a 3.2% improvement. You shipped a 3.2% improvement that happened to show 12% during your experiment window." 3. AI has changed what he's willing to test. Andrea uses AI to turn plain-English test descriptions into production-ready JavaScript variations in under a minute. "AI didn't just save me time. It changed what we're willing to test." The shift he sees coming: from significance-seeking to decision-policy thinking. Instead of "is this significant?" the question becomes "given our traffic, our typical effect sizes, and our tolerance for each error type, what decision rules should we enforce?"

We measured the real A/B testing payload across 8 platforms. The advertised numbers don't match what actually loads. We set out to answer one question: what's the actual payload required to run an A/B test in production? The measurement covered the full execution footprint, not just the advertised snippet. Every asset required to deliver an experiment was included: scripts loaded at runtime, total bytes transferred, network requests triggered, and dynamic loading patterns. Eight platforms were measured: Convert, ABlyft, VWO, Mida, Webtrends, Visually, Fibr, and Amplitude. For four of them, tracking scripts were pulled directly from live customer sites or public GitHub repositories, then measured with curl. The rest were validated against the Mida benchmark and vendor documentation. What the research found is that the advertised snippet size rarely reflects the total payload. Vendors report the initial loader. The actual execution cost sits behind additional scripts, API fetches, and dynamically injected resources that only show up in a full network waterfall. The gap matters because Core Web Vitals, flicker prevention, and data reliability all depend on the real payload, not the marketed one. Full breakdown of how Convert's delivery architecture differs from every platform measured - link in the comments.

Welcoming Schemerr to the Convert partner network. Imagery-first CRO for Shopify brands. Schemerr is a founder-led CRO agency built for Shopify brands. Dayo Samuel runs point on every account, with a hand-picked senior squad behind him: strategist, designer, engineer, director. The approach is called Imagery-First CRO. It starts with category-correct AI product images, shipped in 48 hours through their visual intelligence engine (AdaptiveIQ). Once the imagery lift is measured, the CRO engine takes over: hypothesis, design, test, implement, repeat. Ten categories, ten different conversion playbooks. The squad embeds in your stack so testing runs at in-house speed, without the headcount. Recent work includes Source, Step One Foods, Tradewind Group, and eManuals. They also run a free PDP teardown: drop your product page URL, get a one-page score with your three biggest conversion issues. No call required. Welcome to the certified partner network, Schemerr!

AI is in every CRO workflow now. 5 experts told us how it changed theirs. We asked five experimentation pros the same question: How is AI influencing your experimentation workflows? Their answers came from different angles but landed in the same place. Pieter Boonstra sees AI handling everything around the work: experiment setup, analyses, coding, statistics. The point is to free people up for what only humans can do, which is coming up with solutions for other people. Matthew Bass uses it to compress days of analysis into hours. AI synthesises qualitative and quantitative data, pressure-tests ideas, and surfaces untapped hypotheses. "It increases the leverage of leaders who already know how to ask the right questions." Dzifa Mensah calls it a "reliable thinking partner in the pre-test phase." She also uses it for stakeholder management: drafting test announcements, writing readable reports, and communicating results to non-technical audiences. Katie Faulkner treats AI like a team member. "While the agent is doing that, we'll be doing it too. We meet halfway." Her sharpest line: "AI should be on that list of minds. But that's the important word: one of those minds, not the mind." Bruno Borges uses AI most for research, especially organising large customer survey datasets and generating hypotheses. He's tested identical prompts across multiple AI tools and Claude consistently gives him the most accurate, structured insights. The same pattern shows up in every interview: AI handles the synthesis, humans handle the thinking. Whoever knows the right question to ask still leads.

You can win on the landing page and lose by day 7. The first week decides if they stay. Daphne Tideman wrote an article making the case that the first 7 days after a customer signs up or buys is the most underrated CRO opportunity. The argument: most teams optimise the landing page, the checkout, the form. The biggest drop-offs happen after the conversion. Her framework for fixing it: 1. Define what "activated" means for your product. Pull two groups from your data: customers who churned within two weeks, and customers still active after 60 days. Look at what they did differently in week one. 2. Audit your post-purchase experience as if you were a new customer. Buy the product, sign up, go through every email and prompt. For each moment, ask: is this building trust, creating clarity, increasing motivation, or just noise? 3. Experiment on the moments that matter most. For most businesses, day 0 (the moment of purchase) and days 1-3 (first use) are where the biggest drops happen. 4. Don't forget pre-purchase messaging. What you communicate before someone buys shapes how they behave afterward. What to measure: day 1 and day 7 retention, time to first key action, and second session rate. Conversion rate alone misses the point. The compounding effect Daphne flags is the bit worth saving. A 10% lift in landing page conversion gives you 10% more customers. The same lift in 7-day retention means those customers stay longer, spend more, and refer others. One adds. The other multiplies. convert.com/blog/growth-ma…

21 lessons that walk you from Ikigai to your first published post. Posting isn't step one. Personal Branding for Women in Growth & Experimentation is the cohort Lucia van den Brink and Daphne Tideman are running starting July 10. The course unfolds over three weeks: - Positioning (Week 1). You work through the Ikigai framework, identify three content pillars, choose your channel, and map out a monetisation roadmap. - Profile (Week 2). You rewrite your headline (220 characters, three questions), redesign your banner, restructure your About section, and build an engagement routine that creates relationships instead of just impressions. - Content system (Week 3). You learn what good posts do, write hooks, find your tone of voice, edit drafts, and develop the habit of posting through writer's block. By the end, you walk out with a positioning statement, a profile rewrite, a content system, and a published post. Lucia runs The Initial, a CRO consultancy working with British Airways, Semrush, and Independer, and has 17K LinkedIn followers. Daphne built a fractional growth practice entirely on inbound and has 26K LinkedIn followers. They built both alongside the work, not instead of it. One more detail worth flagging: for every paid ticket, a free spot opens for someone in the community who's out of work or in financial difficulty. Buying a place gives one too. Cohort runs July 10-24, 2026. Course outline and how to enroll: maven.com/women-in-exper…

"The silent killers of good A/B tests." That's the first section of Deborah O'Malley's new course. The five hours is worth it. Deborah is the founder of GuessTheTest and a guest speaker at the Nielsen Norman Group. Her new course is called A/B Testing: A Complete Guide. The structure: six sections, 36 videos, five hours of content, plus practice activities. What it covers: - Silent killers of good A/B tests (nine lessons on what quietly breaks test reliability) - How to choose impactful design ideas to test - Determining sample size and testing duration - Choosing the right testing platform - Analyzing and presenting results stakeholders trust The framework is built on years of analysing tests across hundreds of brands for GuessTheTest. It covers the rigorous side: how to identify wins, losses, and inconclusive results, and how to communicate each one without overstating what the data actually shows. Geared at intermediate-to-advanced practitioners. Worth it if you're running A/B tests at scale or want to learn how to spot suspicious results in tests other teams ran. nngroup.com/contents/self-…