The push layer returns: Why ‘publish and wait’ is half a strategy

In 1998, submitting a website to search engines was manual, methodical, and genuinely tedious. I remember 17 of them: AltaVista, Yahoo Directory, Excite, Infoseek, Lycos, WebCrawler, HotBot, Northern Light, Ask Jeeves, DMOZ, Snap, LookSmart, GoTo.com, AllTheWeb, Inktomi, iWon, and About.com.

Each had its own form, process, and wait time, and its own quiet judgment about whether your URL was worth including. We submitted manually, 18,000 pages in all. Yawn.

Google was barely a year old when we were doing this. But they were already building the thing that would make submission irrelevant.

PageRank meant Google followed links, and a site that other sites linked to would be found whether it submitted or not. The other 17 engines waited to be told about content. Google went looking, and within a few years, they got so good at finding content that manual submission became the exception rather than the norm.

You published, you waited, the bots arrived. For 20 years, that was the deal, and SEO optimized for a crawler that would show up sooner or later.

The irony is that we’re now shifting back. Not because Google got worse at finding things, but because the game has expanded in ways that pull alone can’t cover, and the revenue flowing through assistive and agentic channels doesn’t wait for a bot.

Pull isn’t the only entry mode

The pull model (bot discovers, selects, and fetches) remains the dominant entry mode for the web index. What’s changed is that pull is now one of five entry modes into the AI engine pipeline (the 10-gate sequence through which content passes before any AI system can recommend it), not the only one. 

The pipeline has expanded, and new modes have been added alongside the existing model rather than replacing it, and the single entry mode that has been the norm for 20 years has expanded to five.

What follows is my taxonomy of those five modes, with an explanation of the advantages each one gives you at the two gates that determine whether content can compete: indexing and annotation.

The five entry modes differ by gates skipped, signal preserved, and revenue reached

Mode 1: Pull model

Traditional crawl-based discovery where all 10 pipeline gates apply and the bot decides everything. You start at gate zero and have no structural advantage by the time your content gets to annotation (which is where that content starts to contribute to your AI assistive agent/engine strategy). You’re entirely dependent on the bot’s schedule and the quality of what it finds when it arrives.

Mode 2: Push Discovery

The brand proactively notifies the system that content exists or has changed, through IndexNow or manual submission. 

Fabrice Canel built IndexNow at Bing for exactly this purpose: “IndexNow is all about knowing ‘now.’” It skips discovery, improves the chances of selection, and gets you straight to crawl. The content still needs to be crawled, rendered, and indexed, because IndexNow is a hint, not a guarantee. 

You win speed and priority queue position, which means your content is eligible for recommendation days or weeks earlier than a competitor who waited for the bot. In fast-moving categories, that window is the difference between being in the answer and being absent from it.

Note: WebMCP helps with Modes 1 and 2 by making crawling, rendering, and indexing more reliable, retaining signal and confidence that would otherwise be lost through those three gates. 

Because confidence is multiplicative across the pipeline, a higher passage rate at crawling, rendering, and indexing means your content arrives at annotation with significantly more surviving signal than a standard crawl delivers. The structural advantage compounds from there.

Mode 3: Push data 

Structured data goes directly into the system’s index, bypassing the entire bot phase. Google Merchant Center pushes product data with GTINs, prices, availability, and structured attributes. OpenAI’s Product Feed Specification powers ChatGPT Shopping that supports 15-minute refresh cycles. 

Discovery, selection, crawling, and rendering don’t exist for this content, and the “translation” at the indexing phase is seamless: it arrives at indexing already in machine-readable format, four gates skipped and one improved. That means the annotation advantage is significant.

This is where the money is for product-led businesses: where crawled content arrives as unstructured prose the system has to interpret and feed content arrives pre-labeled with explicit machine-readable entity type, category, and attributes. By structuring the data and injecting directly into indexing, you’re solving a huge chunk of the classification problem at annotation, which, as you’ll see in the next article, is the single most important step in the 10-gate sequence.

As the confidence pipeline shows, each gate that passes at higher confidence compounds multiplicatively, so this is where you can get the “3x surviving-signal advantage” I outline in “The five infrastructure gates behind crawl, render, and index.”

Mode 4: Push via MCP 

Model Context Protocol (MCP) — a standard that lets AI agents query a brand’s live data during response generation — allows agents to retrieve data from brand systems on demand. 

In February 2026, four infrastructure companies shipped agent commerce systems simultaneously. Stripe, Coinbase, Cloudflare, and OpenAI collectively wired a real-time transactional layer into the agent pipeline, live with Etsy and 1 million Shopify merchants. 

Agentic commerce is key. MCP skips the entire DSCRI pipeline and then operates at three levels, each entering the pipeline at a different gate: 

• As a data source at recruitment.
• As a grounding source at grounding.
• As an action capability at won, where the transaction completes without a human in the loop. 

The revenue consequences are already real: brands without MCP-ready data are losing transactions to those with it, because the agent can’t access their inventory, pricing, or availability in real time when it needs to make a decision. This is where you see multi-hundred percent gains in the surviving signal.

MCP is already simultaneously push and pull, depending on context. 

There’s a dimension to Mode 4 that most people don’t think about much: the agent querying your MCP connection isn’t always a Big Tech recommendation system. It’s increasingly the customer’s own AI, acting as their purchasing agent, evaluating your inventory and pricing in real time, with their credit card behind the query, completing the transaction without them opening a browser.

When your customer’s agent (let’s say OpenClaw-driven) comes knocking, agent-readable is the entry requirement. Agent-writable — the capacity for an agent to act, not just retrieve — is where you’ll make the conversion. The brands without writable infrastructure will be losing transactions to competitors whose systems answered the query and handled the action.

Mode 5: Ambient

This is structurally different from the other four. Where Modes 1 through 4 change how content enters the pipeline, ambient research changes what triggers execution of the final gates. 

The AI proactively pushes a recommendation into the user’s workflow without any query: Gemini suggesting a consultant in Google Sheets, a meeting summary in Microsoft Teams surfacing an expert, and autocomplete recommending your brand. 

Ambient is the reward for reaching recruitment with accumulated confidence high enough that the system fires the execution gates on the user’s behalf, without being asked. You can’t optimize for ambient directly. You earn it — and the brands that earn it capture the 95% of the market that isn’t actively searching.

Several people have told me my obsession with ambient is misplaced, theoretical, and not a real thing in 2026. I’ve experienced it myself already, but the clearest demonstration came at an Entrepreneurs’ Organization event where I was co-presenting with a French Microsoft AI specialist. 

He demonstrated on Teams an unprompted push recommendation: a provider identified as the best solution to a problem his team had been discussing in the meeting. Nobody explicitly asked. Copilot listened, understood the problem, evaluated options, and push-recommended a supplier right after the meeting. Ambient isn’t theoretical. It’s running on Teams, Gmail, and other tools we all use daily, right now.

---

Every mode converges at annotation

Five entry modes, each with a different starting point, and they all converge at annotation. Annotation is the key to the entire pipeline. Every algorithm in the algorithmic trinity (LLM + knowledge graph + search) doesn’t use the content itself to recruit, it uses the annotations on your chunked content, and nothing reaches a user without being recruited. 

Why is that important? Because accurate, complete, and confident annotation drives recruitment, and recruitment is competitive regardless of how content entered. A product feed arriving at indexing with zero lost signal competes at recruitment with a huge advantage over every crawled page, every other feed, and every MCP-connected competitor that entered by a different door. 

You control more of this competition than most practitioners assume, but skipping gates gives you a structural advantage in surviving signal. It doesn’t exempt you from the competition itself.

That distinction matters here because annotation sits at the boundary. It’s the last absolute gate: the system classifies your content based on your signals, independently of what any competitor has done. Nobody else’s data changes how your entity is annotated. That makes annotation the last moment in the pipeline where you have the field entirely to yourself.

From recruitment onward, everything is relative. The field opens, every brand that passed annotation enters the same competitive pool, and the advantage you carried through the absolute phase becomes your starting position in a winner-takes-all race. Get annotation right, and you have a significant head start. Get it wrong, and no matter how much work you do to improve recruitment, grounding, or display, it will not catch up, because the misclassification and loss of confidence compound through every gate downstream.

Nobody in the industry was talking about this in 2020. I started making the point then, after a conversation on the record with Canel, and it still isn’t getting the attention it deserves.

Annotation is your last chance before competition arrives.

Search is one of three ways users encounter brands — and it’s the least valuable

The research modes on the user’s side have expanded, too. The SEO industry has traditionally focused on just one: implicit, when the user types a query. There was always one more: explicit brand queries, and now we have a third. Each research mode is defined by who initiates and what the user already knows.

Explicit research is the deliberate query, where the user asks for a specific brand, person, or product, and the system returns a full entity response (the AI résumé that replaces the brand SERP). 

This is the lowest-confidence mode of the three, because the user has already signaled very explicit intent: you’re only reaching people who already know your name. Bottom of the funnel, decision. Algorithmic confidence is important here to remove hedging (“they say on their website,” “they claim to be…”) and replace it with absolute enthusiasm (“world leader in…,” “renowned for…”).

Implicit research removes the explicit query. The AI introduces the brand as a recommendation (or advocates for you) within a broader answer, and the user discovers the brand because the system considers it relevant to the conversation, staking its own credibility on the inclusion. Top- and mid-funnel, awareness and consideration. Algorithmic confidence is vital here to beat the competition and get onto the list when a user asks “best X in Y market” or be cited when a user asks “explain topic X.”

Ambient research requires the highest confidence of all. The system pushes the brand into the user’s workflow with no query, no explicit request, the algorithm is making a unilateral decision that this user, in this context, at this moment, needs to see your brand. That requires very significant levels of algorithmic confidence.

The format is small: a sentence, a credential, a contextual mention. The audience reached is the largest: people not yet in-market, not yet actively looking, who encounter your brand because the AI decided they should. And the kicker is that your brand gets the sale before the competition even starts.

For me, this is the structural insight that inverts how most brands prioritize, and where the real money is hiding. They optimize for implicit research, where competition is highest, the target you need to hit is widest, and the work is hardest. 

Most SEOs underestimate explicit research (where profitability is highest) and completely ignore ambient, which reaches the 95% who aren’t yet looking and requires the deepest entity foundation to trigger. I call this the confidence inversion, first documented in May 2025: the smallest format requires the highest investment, and it reaches the most valuable audience.

The entity home website is the single source that feeds every mode

In 2019, AI engineers spent 80% to 90% of their time collecting, cleaning, and labeling data, and the remaining 10% to 20% on the work they actually wanted to do. They wryly called themselves data janitors. Today, Gartner estimates 60% of enterprises are still effectively stuck in the 2019 model, manually scrubbing data, while the companies that got organized early compound their advantage.

The same split is happening with brand content and entity management, for the same reason. Every push mode described in this article draws on data: product attributes for merchant feeds, structured entity data for MCP connections, and corroborated identity claims for ambient triggering. 

If that data lives in scattered, inconsistent, contradictory sources, every push attempt is expensive to implement, structurally weak on arrival, and liable to contradict the previous one. Inconsistency is the annotation killer: the system encounters two different versions of who you are from two different push moments, and confidence drops accordingly.

The framing gap, where your proof exists but the algorithm can’t connect it to a coherent entity model, is a direct consequence of disorganized data, and it costs you in recommendation frequency every day it persists.

The entity home website — the full site structured as an education hub for algorithms, bots, and humans simultaneously, built around entity pillar pages that declare specific identity facets — becomes the single source that feeds every mode simultaneously.

Pull, push discovery, push data, MCP, and ambient all draw from the same clean, consistent, non-contradictory data. You build the structure once, maintain it in one place, and you’re ready for push and pull modes today, and any to come that don’t yet exist.

AI handles 80%, humans protect the other 20%

That foundation is only as strong as the corrections made to it. How this works in practice depends on where you’re starting from. For enterprises, the website typically mirrors an internal data structure that already exists: 

• Product catalogs. 
• CRM records.
• Service definitions.
• Organizational hierarchies. 

The website becomes the public representation of structured data that lives inside the business, and the primary challenge is integration and maintenance.

For smaller businesses and personal brands, the direction often runs the other way: building the entity home website well is what forces you to figure out how your business is actually structured, what you genuinely offer, who you serve, and how everything connects. The website imposes discipline. 

We’re doing exactly this: centralizing everything as the structured data representation of the entire brand (personal or corporate). Getting the foundation right (who we are, what we offer, who we serve) is generally the heaviest lift. Building N-E-E-A-T-T credibility on top of that foundation is now comparatively straightforward, and every new push mode draws from the same organized source.

Here’s where using AI fits into this work. It can handle roughly 80% of the organization: extracting structure from existing content, proposing taxonomies, drafting entity descriptions, mapping relationships, and flagging gaps. What it does poorly, and what humans need to correct, are the three failure modes that propagate silently through every downstream gate:

• Factual errors, where something is simply wrong.
• Inaccuracies, where something is approximately right but imprecise enough to mislead.
• Confusions, where two different concepts are conflated, or an entity is ambiguous between interpretations.

Confusion is the sneakiest because it looks like data, passes automated quality checks, enters the pipeline with apparent confidence, and then causes annotation to misclassify in ways that compound through every gate downstream.

Alongside the errors sit the missed opportunities, which are equally costly and considerably less obvious:

• Lost N-E-E-A-T-T credibility opportunities, where the systems underestimate or undervalue the entity because credibility signals exist but aren’t structured, corroborated, or framed in a way the algorithmic trinity can read. The authority exists, but the machine doesn’t understand it.
• Annotation misclassification, where the entity is indexed coherently but placed in the wrong category, meaning it competes for the wrong queries entirely and never appears in the contexts where it should win. Correctly classified competitors take the recommendation: your brand is present in the pipeline, but absent from the competition that matters to your business.
• Untriggered deliverability, where understandability is solid and credibility has crossed the trust threshold, but topical authority signals haven’t accumulated densely enough to push the entity across the deliverability threshold for proactive recommendation. The machine knows who you are and trusts you. It just doesn’t advocate for you yet.

The human doing the correction and optimization work is the competitive advantage. Because the errors are surreptitious and the opportunities non-obvious, the trick is finding where both actually are, fixing one, and acting on the other.

The errors are surreptitious. The opportunities are non-obvious. Finding both is the work that compounds.

Organize once, feed every mode that exists and every mode to come

The push layer is expanding. The brands that organize their data now — not perfectly, but consistently, and with a system for maintaining it — are building the infrastructure from which every current and future entry mode draws.

The brands still publishing and waiting for the bot (Mode 1) are optimizing for the least advantageous mode in a five-mode landscape, and that disadvantage gap widens with every passing cycle.

This is the seventh piece in my AI authority series. 

• The first, “Rand Fishkin proved AI recommendations are inconsistent – here’s why and how to fix it,” introduced cascading confidence. 
• The second, “AAO: Why assistive agent optimization is the next evolution of SEO,” named the discipline. 
• The third, “The AI engine pipeline: 10 gates that decide whether you win the recommendation,” mapped the full pipeline. 
• The fourth, “The five infrastructure gates behind crawl, render, and index,” walked through the infrastructure phase.
• The fifth, “5 competitive gates hidden inside ‘rank and display’,” covered the competitive phase.
• The sixth, “The entity home: The page that shapes how search, AI, and users see your brand,” mapped the raw material.
• Up next, “How AI decides what your content means (and why it gets you wrong).”

View the full article