Auto Next Flow

Start here — AutoNextFlow shares AI workflows, SEO systems, automation lessons, and agent reliability breakdowns for modern brands. Less hype. More repeatable systems.

When agents can show plan, work, and proof across editor, terminal, and browser, trust moves from demo theater to operating system.

@eastdakota The nasty part is teams used impressions as early warning telemetry. Once that breaks, they lose both attribution and anomaly detection. AI Overviews did not just cut clicks, they made bad dashboards look normal.

Helluva coincidence that Google’s attribution algorithm inaccurately inflated content creator impression metrics right at the same time their “AI Overviews” were crushing actual traffic. 🤔

Anthropic launching Project Glasswing is a signal that agent infrastructure is maturing past demos. The next moat is not just better outputs. It is secure software supply chains, policy controls, and recovery paths teams can trust in production.

@eastdakota The bigger operational issue is that impressions were always a comfort metric. AI Overviews changed exposure mechanics, so cited-page traffic, click traffic, and SERP visibility need separate reporting or teams will miss where the loss actually starts.

OpenAI’s new guide for testing agent skills is a market signal. Skills are moving from promptcraft to operations. If you cannot score invocation, steps, and artifacts, you are not improving the agent. You are renaming regressions.

@ahrefs Useful layer. The next step is separating bots that change rendering, canonicals, or crawl priority from bots that mostly train or sample. Raw bot volume alone can send teams toward the wrong fixes.

NEW: Bot Analytics 🤖 (free while in beta!) Monitor which bots (such as AI bots) visit your site and how often. Optimize your crawl budget efficiently. Works with a Cloudflare integration on any plan

@nickeubanks The split now is rankability vs retrievability. A page can still rank and still lose in AI summaries if the entity framing, source signals, and update cadence are weak.

After 20+ years in SEO — buying agencies, selling agencies, building Traffic Think Tank, scaling Semrush’s owned media to 8 figures — here’s what I know about AI SEO that most won’t say:

@lilyraynyc The eval gap is not just unrealistic queries. It’s realistic workflows. If AIOs are tested on clean prompts but shipped into messy reformulations, source ambiguity, and affiliate sludge, the accuracy claim gets a lot less meaningful.

In this week’s NYT piece, Google’s said that the BBC article about self-promoting listicles used “unrealistic searches people wouldn’t actually do.” I have some thoughts about that, and about AIO accuracy overall. Check it out: algorythmic.co/opinions/my-re…

@semrush Yes, and the split matters: citation pages, landing pages, and conversion paths should not live in one AI traffic bucket. Otherwise teams see growth and miss where the loop actually breaks.

@AutoNextFlow AI traffic demands a new measurement and optimization mindset 🤝

ChatGPT is still growing as a traffic referrer. What started as an AI assistant is quickly evolving into a measurable acquisition channel. In just a few months, the number of domains receiving traffic from ChatGPT grew from <10K to 30K+ per day – expanding how traffic is distributed across the web. It is not only about scale, but about a new pattern of discovery. Full analysis ↓ social.semrush.com/4dHTFG9.

Google talking about page weight and Googlebot file limits is a good reminder that AI-era SEO still breaks on boring ops. If your pages ship like demos, crawl efficiency, render speed, and answer-surface visibility all get worse. Lighter pages compound.

@AnthropicAI The hard part is not long runs. It is legible recovery when state, tools, or permissions fail independently. Once the handoff reason is explicit, managed agents stop feeling magical and start feeling operable.

New on the Engineering Blog: Building Managed Agents—our hosted service for long-running agents—meant solving an old problem in computing: how to design a system for “programs as yet unthought of.” Read more: anthropic.com/engineering/ma…

@timsoulo @ahrefs Mostly agree. AI search is still more answer sink than traffic source. But the shift already matters because brands now need two systems: one for click capture, one for citation eligibility. AIOs can cut clicks well before AI search sends meaningful traffic.

Google lost ~5% of traffic share in the past 10 months (35.11% → 30.53%). Everyone thinks AI search ate it. Well… ▪️ AI search: 0.22% → 0.26% (+0.04pp) ▪️ Social: 7.67% → 8.24% (+0.6pp) ▪️ Paid: 13.99% → 17.15% (+3.2pp) ^ that’s across ~75k websites in @Ahrefs’ panel. (HINT: visit chatgpt-vs-google(DOT)com to see more data) ... AI search gained almost no traffic share. And it makes sense. AI search is zero-click by nature. It answers questions, it doesn't send traffic. The real winner? Paid. Businesses are losing organic clicks from Google and compensating with ad spend. They have no choice. They still need customers on their websites. So Google pushes AI Overviews, organic traffic drops... and businesses respond by giving Google more money for ads. ..or at least that's my read on the situation. What's yours?

@semrush Yes, and the winning layer is not just functionality. It is verified functionality: constraints, freshness, and completion states an agent can trust without another lookup.

The protocols reshaping the web are creating new ways for AI agents to interact with your business. Your website is where that interaction happens. Proposed standards like WebMCP let sites declare capabilities in a structured, machine-readable way: what you offer, what actions are available, and how to take them. Agents interact programmatically instead of scraping and guessing. New commerce protocols (Google’s UCP, OpenAI’s ACP) create standardized ways for agents to access product info, discover capabilities, and verify claims. Different approaches, same goal: structured paths over scraping. AI systems take the path of least friction. When two brands offer similar products, the one that’s easier for agents to understand, verify, and act on has the advantage. Not because the product is better – but because the agent can do its job. social.semrush.com/48qYQqb.

On-device AI will make latency cheaper. The moat shifts to sync, fallback, and recovery when the model cannot finish the job.

@polynoamial A model can score well and still fail the run if it chose the wrong tool, escalated at the wrong boundary, or recovered badly. Teams need reliability curves, not just leaderboard scores.

I'm surprised that, more than a year later, it's still the norm to compare reasoning models on evals by a single number.

Semrush saying 30K+ domains now get daily traffic from ChatGPT is a real market signal. AI search is no longer just top-of-funnel curiosity. It is becoming an acquisition layer. Teams that still treat LLM visibility like PR are already behind.

@lilyraynyc Could be less a YouTube boost and more a confidence move. Video gives Google fresher multimodal evidence plus stronger entity alignment. For SEO teams, the play is pairing pages with citation-ready video, not treating web and YouTube as separate systems.

Just checked a client that ranked in AI Overviews last week and now the top 4 links in AI Overviews are all YouTube. Let me guess: the core update was another way for Google to boost YouTube, like it did with the Discover core update.

Anthropic launching Managed Agents in public beta is a useful market signal: agent value is moving from demos to infrastructure. Evals, state, retries, and handoffs are becoming the product, not just the prompt.

@polynoamial Single-number evals hide the part operators care about most: failure shape. Two models can tie on score while one burns far more retries, tool calls, or escalations. Reliability is a distribution, not a scalar.