Growth Forks

The only account you need for growth & marketing

[Alex Groberman @alexgroberman]

Google accidentally revealed how its AI search systems choose which sites to recommend and send traffic to.

Now that none of it is much of a secret anymore, let’s talk about it.

With the new Google Search rolling out as we speak, it has never been more important to understand how to maximize value from this marketing channel.

If you want to see where your site stands across Google, ChatGPT, Claude, Perplexity and broader AI search, you can do so for free here:

https://t.co/Pn764BHwyL

Let’s start from the beginning.

Metehan Yesilyurt, who previously went viral for his analysis of Perplexity’s ranking factors, recently broke down Google AI ranking factors in a fascinating blog post.

A lot of what he found also lines up with what SEO Stuff (https://t.co/wKpf0EILTx) has been doing all year to get customers more traffic and sales.

Specifically with the done-for-you package:

https://t.co/yEFyM0Ze7W

Basically, Google sells much of the underlying retrieval infrastructure through Google Cloud Discovery Engine, also called Vertex AI Search.

By analyzing what Discovery Engine exposes, Yesilyurt was able to show how Google’s AI retrieval and ranking stack works.

The seven major signals he identified were Base Ranking, Gecko embedding similarity, Jetstream cross-attention relevance, BM25 keyword matching, predicted click-through rate, freshness and Boost/Bury rules.

In simpler terms, Google is combining traditional ranking signals with semantic similarity, deeper contextual relevance, keyword matching, engagement predictions, recency and manual business logic.

Discovery Engine exposes a maximum chunk size of 500 tokens, roughly 375 words, while ancestor headings can travel with each chunk and tables, images and page layouts can all be parsed with Gemini-enhanced understanding.

That means important ideas need to make sense inside relatively small, clearly labeled sections.

If your content is one massive wall of text, Google has a much harder job extracting the exact information it needs.

This is also why I keep talking about question-based headings, direct answers, clear comparisons and structured pages.

Additionally, Discovery Engine shows Google processing structured data using separate Searchable, Indexable and Retrievable settings.

Those can affect whether information can be found, used for filtering or ordering and surfaced by the model. Mind you, they are independent.

So a field can potentially influence retrieval or ranking without necessarily being displayed in exactly the same way.

That gives us a pretty significant clue about why structured data continues to matter as Google becomes more AI-driven.

Discovery Engine also reveals a four-stage pipeline: Prepare, Retrieve, Signal and Serve.

Google first interprets the query, handles things like synonyms and intent, then retrieves and chunks relevant content, applies ranking signals and finally uses Gemini to generate the response with grounding and safety rules.

Traditional Search, AI Overviews and AI Mode can be thought of as different configurations built on much of the same underlying retrieval infrastructure.

So what does all of this mean for businesses?

You need to optimize for several layers at once.

Your content needs strong semantic alignment so Google clearly understands what the page is about, enough contextual detail to handle nuanced searches such as comparisons and “best for” queries, and clean chunk-level structure so useful answers can be extracted without digging through hundreds of words of filler.

That means clear definitions, direct answers, comparison language, question-based headings, concise sections, factual claims, TLDR summaries, clean HTML and useful tables where appropriate.

This is exactly why SEO Stuff content is structured the way it is.

The done-for-you package combines 10 long-form, search-focused articles with three DR50+ contextual backlinks:

https://t.co/yEFyM0Ze7W

The content is designed around semantic relevance, keyword alignment, structured sections, comparisons, freshness and extractability, while the backlinks help strengthen the underlying authority of the pages and domain.

The Premium Content Bundle takes the same idea much further:

https://t.co/4CAnUt07PO

It includes 60 long-form articles designed to build broader category coverage, stronger semantic associations and more opportunities for Google to retrieve useful sections across the different questions customers ask.

Then there is the Premium Backlink Bundle:

https://t.co/Z9m9D7TjES

It adds three contextual DR50+ backlinks designed to strengthen search authority and reinforce the company’s identity and category across the web.

And if you want to see where your site stands across Google, ChatGPT, Claude, Perplexity and broader AI search, you can do so for free here:

https://t.co/Pn764BHwyL

[Growth Forks @GrowthForks]

@kanavtwt @dhh i'd buy it if suspend/resume and sleep drain were genuinely boring

curious: would you optimize for framework-style repairability or macbook-style fit and finish?

[Growth Forks @GrowthForks]

@djcows interesting constraint isn't range, it's deployment.

how do you keep the array stable and productive once wind and boat speed stop agreeing?

[Growth Forks @GrowthForks]

@ErnestoSOFTWARE love how boring the answer becomes once you've done it a few times.

[Growth Forks @GrowthForks]

@MichLieben everyone assumes signal quality is the bottleneck. most of the time, the bottleneck is how fast the team acts before the signal goes stale.

[Growth Forks @GrowthForks]

@heyblake the painful part is realizing your customers wrote the headline months ago already

[Growth Forks @GrowthForks]

@denohawari some of the best ideas we've found didn't come from complaints. they came from weird workarounds people were using and proud of.

[Growth Forks @GrowthForks]

@coldemailchris agree, although content gets much easier once you've had enough customer conversations to know what to say.

[Growth Forks @GrowthForks]

@iamfra5er yeah, we've seen this with model swaps too. small quality changes add up fast when the product is customer-facing.

[Growth Forks @GrowthForks]

@IdrisEcom_email yep. a lot of retention problems start earlier than people think. you usually win the second purchase in the first week.

[Growth Forks @GrowthForks]

@atishayhyperke what changes most in this engine when you go from local service businesses to enterprise software?

[Growth Forks @GrowthForks]

@manucastilloll still wild to me how often boring utility wins

[Growth Forks @GrowthForks]

@MichaelSaruggia we've seen the same. getting the list is easy now. keeping deliverability healthy is still where most outbound succeeds or fails.

[Growth Forks @GrowthForks]

@codyschneider curious where the agent still needs the most human input today: audience, offer, or creative direction?

[Growth Forks @GrowthForks]

@jacobrodri_ what's the thing you see most often in the best onboarding flows?

[Atishay @atishayhyperke]

Why do so many cold emailers use AI?

A cold email is ~70-80 words.

It takes 1-5 minutes.

Quicker than writing a prompt 😂

Not to mention

Writing copy with AI is one of the WORST things you could do for performance.

Just makes no sense to me...

Not only does AI copy come out wonky…

But it mangles a chunk of your sends WITHOUT telling you.

And you only find out when a prospect replies:

“F*ck off SPAMMER”

By then that domain has already taken the hit.

Point AI at the grunt work instead:

• reading your replies
• mining the data.

Trust me when I tell you:

The 70 words in the email still belong to a human.

[Atishay @atishayhyperke]

We booked 2,000 meetings in 30 days for our clients...

Using this 3-legged outbound engine (my system, for free):

For context:

Offer is almost never the problem with an underperformign cold email campaign.

It’s usually the “engine” behind it.

An outbound engine is a 3-legged stool:

1. tech
2. strategy
3. follow-up.

Usually just one leg is strong.

• Great infrastructure but a weak offer
• Sharp offer but bad appointment setting

Etc, etc.

Pull any one leg and the whole thing falls → your calendar goes empty.

So I laid out the entire engine we use in one doc.

What is inside:

1) The meeting-math calculator. Reverse from your meeting target to the exact number of interested replies and cold emails you need, using the 25 to 33% positive-reply-to-meeting rate.
2) The 3-legged-stool build spec. Tech, strategy, and follow-up laid out as one engine, so you can see which of the 3 legs is actually missing.
3) Pillar 1, the sending layer. The per-inbox send ceilings by provider, the warm-up ratios, and the 90-day domain-age rule that keep a 7M-emails-a-month operation out of spam.
4) Pillar 2, the testing loop. The offer-angle-audience triplet, the 4 WHYs every cold email has to answer, and the 1,000-prospect minimum test size before you can trust the result.
5) Pillar 3, the conversion play. The 1-minute response rule, the Acknowledge-Associate-Ask reply framework, and the 8-to-12-touch subsequence that runs over 3 weeks.
6) The "which leg is broken" debug tree. Match your broken metric (total reply rate with out-of-office counted, positive replies, or booked meetings) to the exact pillar to fix, in order.

PS. The same engine has held up across 150+ industries, from local service shops to enterprise software.

[Growth Forks @GrowthForks]

@OriSilver curious how often the analyst flags things users don't actually care about. that's usually the gap we run into.

[Growth Forks @GrowthForks]

@helloitsolly clicking through to the original post hits a very specific trust itch for me

[Growth Forks @GrowthForks]

@Quentin_Rnv most of this feels obvious until you end up paying for ignoring it.