Explore

Double down on the best-LTV:CAC channel and hit prospects from many channels

Double down on the acquisition channel with the best LTV:CAC ratio and do way more of it, rather than diversifying channels — stick with LinkedIn where there's already a proven process. Send as many DMs as the platform allows on LinkedIn as the volume channel, hit prospects from as many channels as possible (volume channels will be the constraint before quality channels), call the highest-likelihood responders first because they're warmer while using the dialer for volume, run an AI scrape of contacts to extract phone numbers and feed them into a dialer, and commit to the chosen acquisition activity twice a day, every day.

Alex Hormozi · How to Get Rich for Service-Based Businesses (Thank Me Later)

Use vaults for credential injection in production

In a real production application, use vaults for credential injection: define a secret to inject in an environment variable, put a placeholder instead of the real secret, and at egress swap it out for the real value only on host domains you specified are allowed for that secret. Vaults work like an API: create a vault, give it a name, get an ID, and include that ID in your session as a way to authenticate with your MCPs/CLI. Whenever you reference a session, inject that vault ID and it will bundle all the credentials into your session — whenever it tries to use any of the domains you specify in your credentials, it will swap for the real credential instead of the placeholder.

Method, not a verified protocol.

AI Engineer · Build a Research Agent with Claude Managed Agents (Part 2 of 4) — Anthropic

Implement the event-stream interface to talk to the agent

Tell Claude to continue and implement the interface to talk to the agent in your application. The main way you communicate with a session is through an event stream: listen on that stream for events the agent pushes back, and push events back up through it. When you press enter in the chat box, it sends a user.message event with the content to that session.

Method, not a verified protocol.

AI Engineer · Build a Research Agent with Claude Managed Agents (Part 2 of 4) — Anthropic

Publish best-X roundup articles that include competitors

Create a 'best [category]' roundup article that lists your SaaS alongside competitors, and do it really well. This is the main way to surface in those searches because Google and LLMs prefer these roundups, especially in smaller niche SaaS markets where no affiliate sites fill that role.

Startups For the Rest of Us · Episode 853 | How to Become the SaaS That AI Recommends

Start GEO with transactional content, roundups, and versus articles

The guest recommends that a bootstrap founder start with creating roundups and versus articles — 100% where he would start first — and notes that transactional content on your website is very fruitful for GEO. If you can't immediately identify a directory like Clutch or G2 in your niche, create those roundups yourself and make them high quality.

Startups For the Rest of Us · Episode 853 | How to Become the SaaS That AI Recommends

Position as 'best at a component' rather than 'best overall'

Don't call yourself 'best overall.' Claim a niche 'best for X' position (e.g., best CRM for small business) instead. This lowers friction for third parties to repeat the claim, helps LLMs segment you into a curated summary, and gets you into the conversation.

Startups For the Rest of Us · Episode 853 | How to Become the SaaS That AI Recommends

Prioritize versus articles as an underrated AI-search tactic

Publish 'X versus Y' comparison articles. Siege Media's correlation study across its B2B SaaS client base found versus articles had the highest Spearman correlation with LLM visibility — higher than best-X pages — and sites with the most versus articles had the most overall LLM traffic. G2 data also shows 'X versus Y' searches come shortly after 'best X' searches.

correlational, not causal; publisher-reported

Startups For the Rest of Us · Episode 853 | How to Become the SaaS That AI Recommends

Cross-model prior intersection catches errors: Jev and Luna agreed on 194 of 200, all errors in disagreements

You can get really far with cross-model prior intersection: on the 200 held-out families, Jev and Luna agreed on 194, and none of those shared answers was wrong; every error sat in the six families where they disagreed.

AI Engineer · Just-in-Time Ontologies: What Cheap Intelligence Means for Data — Hrishi Olickel, Southbridge AI

Rewriting questions fixed the failure without touching models or thresholds

Rewriting the questions so the household question stated what counts as enough evidence moved Jev from no on all 13 development questions to yes, without changing models or thresholds.

AI Engineer · Just-in-Time Ontologies: What Cheap Intelligence Means for Data — Hrishi Olickel, Southbridge AI

Get your product mentioned in pages AI already cites — the blueprint for AI Overview visibility

Jake's advice for AI visibility: look up the queries you want to rank for in AI Overviews, reach out to the pages currently being cited, and get your product mentioned there — 'if you can get your product mentioned in pages that AI is already citing, you kind of have the blueprint for that.' When your site is getting cited in AI Overviews, use that visibility data (e.g. via Brand Radar) to identify which pages to prioritize for outreach. The Ahrefs tool pulled a prioritized list of AI citation pages from Brand Radar (101, 62, 56 AI responses). It takes time to actually rank for this stuff, so start early — waiting until you're out of leads and money is tight means you're too late, because when the flow dries up money will be tight. Create a comparison article for every competitor shoe vs.

The comparison-content mechanism (teaching AI systems your product exists) is presented as the host's/Jake's reasoning, not a verified outcome.

Ahrefs · The Greatest Google SEO Recovery of All Time?

Plan the change: assess behavior, ground in repo context, and produce a structured plan

Nnenna Ndukwe's planning phase: switch into plan mode after agent setup is complete, use the repo-local workshop planning skill to turn the hands-on task into a high-level implementation plan and a build session execution plan, and ground the plan in agents.md, README, and the rules. Verify the local Qodo setup and dependency health before selecting rules, and assess the current behavior of the app during planning to determine what needs to be done and the shape of the code change. If the agent doesn't prompt you to choose a task, make that clear as part of the prompt. A plan should answer: what behavior are we changing, which rules apply, what skills would guide the agent, which tests prove the expected behavior, and what quality gates should pass before review.

Endorsed; no stated uncertainty.

Method, not a verified protocol.

AI Engineer · How to Build Quality Gates into Agentic Coding Workflows — Nnenna Ndukwe, Qodo AI

Mitigations and controls: secure pipelines, API protections, and token storage

Define good LLM pipelines with data security, protections for APIs, and rate limiting as mitigations and controls, and store all tokens used by agentic tools in a way that they cannot be leaked.

Method, not a verified protocol.

AI Engineer · AI Security Engineer Foundations + Certificate — Javier Garza, Snyk

Start small: extract structure first, then layer in semantic understanding

To build something actually useful and grounded, start small — start with the smallest possible thing you can do, which is extracting structure out of the codebase. Then move on to layering semantic understanding into that structure to build a substrate useful for agents to work inside the space and make changes.

endorsed

Method, not a verified protocol.

AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef

Parse transformation logic to ASTs and compute facts deterministically

Parse all transformation logic down to the ASTs to understand how the data is actually structured so the agent cannot drift from the actual implementation; compute facts deterministically and check them against your data first, then let the LLM only name the data assets at the end on the graph.

AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef

Tie the password vault to the agent runtime, not the reasoning layer

To keep secrets out of the model, the host recommends tying the password vault to the agentic runtime instead of the reasoning layer; the agent runtime then becomes an exit layer that can be considered deterministic. With this architecture you can do things you normally couldn't for privacy and security reasons at the LLM layer — for example, using a browser-use tool to log into a website with your credentials becomes much more reasonable to call private and secure than feeding your password directly into the reasoning layer.

AI Engineer · How Secrets Leak Through AI Agents (and How to Stop It) — Venice AI

Take the repo and use it to build your IDPs, as a model with Backstage, security rails, and provisioning

The speaker encourages the audience to take the repo and use it to build their IDPs and as a model; if you're looking to stand up an IDP — a platform with Backstage, proper security rails, provisioning, and best practices — this is the one to look at.

AI Engineer · Build a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog

Don't start enterprise MCP adoption with public MCP servers

Forrester discourages starting enterprise MCP adoption by using public MCP servers, because a significant double-digit percentage — he estimates 33-40% — of public MCP servers have horrible security vulnerabilities. He recounts that someone advised an enterprise to just use all the public MCP servers out there as a way to discover MCP, implicitly as a cautionary example.

The 33-40% figure is the speaker's own estimate; flagged possible_numeric_error.

AI Engineer · Build a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog

For enterprise MCP adoption, roll your own, use a vendor, or use a registry/gateway

Forrester recommends that for enterprise MCP adoption, you either roll your own MCP server, get in a relationship with a vendor, or get an MCP registry or a gateway. He notes that most organizations at this level are not wiring MCP servers directly into their platform; they talk to a registry or an MCP gateway that provides a list of allowed tools that get pulled in.

None stated.

AI Engineer · Build a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog

Reducto workflow: create schema, validate on samples, inspect results, then deploy

Omar describes the simple workflow: create a new extract pipeline with a schema for the document type, upload sample documents and run, visually confirm results match expectations (revising the schema if not), then deploy via direct API call or pipeline code snippet — 'stick your documents in this platform and then immediately start visualizing your data, and you could immediately also productionize it afterwards.'

Method, not a verified protocol.

AI Engineer · Parsing the Epstein Files for JMail — Palak Agarwal & Omar Alhait, Reducto

Check token usage and cost before trusting a high benchmark score

Use the 3D token view to check how many tokens a model used to achieve its score before trusting a high benchmark result. Choosing a fast-looking model without checking token usage can backfire: Gemini 3.5 Flash scored nearly as high as GPT 5.5 but used 1.65 billion tokens vs 136 million, costing $620 vs $497. Before assuming a flash model is faster and cheaper, compare token counts and cost — all those tokens also take time. When one model has a lower score but a bigger bill, it makes more sense to use the other model.

AI Engineer · From Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf, Weights & Biases