Ask Jeeves launched in 1996. The proposition was simple and genuinely interesting: type a question in natural language and get a direct answer rather than a list of links. "What is the capital of France?" not "capital France." A butler rather than an index. The name was not accidental.
It failed — not because the idea was wrong but because the technology could not support it. Natural language processing in 1996 was not capable of reliably interpreting intent across the breadth of queries a search engine needs to handle. The index was thin. The answers were often wrong or missing. Google arrived with better retrieval, scaled faster and buried Ask Jeeves under the weight of a superior index. By 2010 the butler had been retired. The service quietly rebranded as Ask.com and eventually became largely irrelevant.
Here is the thing. Google spent the next fifteen years building exactly the infrastructure that Ask Jeeves needed to work. The index. The entity understanding. The natural language processing. The knowledge graph. The ability to interpret intent rather than just match keywords. And now, in 2026, Google is using all of that infrastructure to become exactly what Ask Jeeves was trying to be — a system that answers your question directly, acts on your behalf and increasingly removes the need to visit a website at all.
Jeeves is back. He just works for Google now.
The mechanism: WebMCP and conversational ads together
Two recent developments sit alongside each other in a way that the coverage has not fully connected. Google's announcement of conversational AI inside Search Ads — the Business Agent for leads — and the WebMCP specification are not separate stories. They are the same story told from different ends.
Conversational ads create the interface. A user searches for car insurance, a mortgage, a hotel, a solicitor. Instead of clicking to a landing page they enter a dialogue within the ad. The agent asks qualifying questions. The user responds. The agent, reading your website content via WebMCP's structured annotations, generates answers about your products and services on your behalf.
And then, at the natural end of that conversation: "Would you like me to go ahead and apply on your behalf?"
WebMCP's toolautosubmit attribute is what makes that last
step possible. A boolean that allows an agent to complete and submit a
form without an explicit user confirmation step. The conversation completes.
The form is submitted. The lead is generated — or the quote is requested,
or the appointment is booked — inside Google's interface, using your
website's content, without the user visiting your site.
Nobody clicked anything. There is no click to count.
Cost per click is the wrong metric
CPC as a pricing model assumes a click occurred. If the conversion happens inside the conversational interface — if the user never navigates to your site — the click does not exist. Google will need a new pricing unit for this model and the candidates are all structurally more favourable to Google than to the advertiser.
Cost per conversation: you pay for the dialogue being initiated regardless of outcome. Google controls what constitutes a conversation.
Cost per qualified lead: you pay based on Google's intent scoring. Google controls what constitutes qualification.
Cost per outcome: you pay when the agent completes an action on the user's behalf. Google controls what constitutes completion.
In each case the pricing unit is defined, scored and reported by the platform receiving the payment. CPC was transparent in a way none of these are. One click. One cost. Auditable in your own analytics. The move to outcome-based pricing is presented as better for advertisers because you only pay for results. What it actually means is that you pay for results as Google defines them, using a model Google built, reported through a dashboard Google controls.
The content underneath all of it
The agent's conversation is powered by your website content. WebMCP makes that content machine-readable in a structured way — form fields annotated with descriptions, tools registered with purposes, the site's capabilities declared for agents to understand. But the substance of what the agent says about your products, your pricing, your availability, your differentiators — all of that comes from what your website actually contains.
If your product pages are vague the agent's answers will be vague. If your content describes products generically rather than specifically the conversation will be generic. If your site does not clearly differentiate you from competitors the agent will not differentiate you either. The consequences of weak content just became considerably more serious than a high bounce rate. A bounce means the user left. An agent completing a form with inaccurate information means the user applied for the wrong thing — and may not know it until considerably later.
This is the thread that runs from Ask Jeeves to WebMCP to conversational ads to the question of what your website is actually for. Ask Jeeves failed because the answers were wrong. Google succeeded because the index was right. The agent that completes a booking on a user's behalf is only as good as the content it is reading. Getting that content right is not an SEO task in the traditional sense. It is the foundation of whether the new model works at all.
The measurement problem, compounded
If the conversion happens inside Google's interface the visit to your site never occurs. GA4 records nothing. Your CRM receives a lead from a source that Google has already pre-qualified using signals you cannot inspect. The attribution chain runs: user query — Google agent — WebMCP interaction — form submission — lead delivered. You see the lead. You do not see the chain that produced it.
Google's predictive attribution products are designed to fill this gap — to tell you what the agent's conversation contributed to the eventual outcome. But predictive attribution is a model. The model belongs to Google. The inputs to the model belong to Google. The output is a number that tells you what Google's model decided the interaction was worth.
You just haven't earned the underlying data yet, baby. And the structure being built does not obviously change that.
The organisations that will navigate this most effectively are the ones that maintain independent measurement alongside Google's reporting — their own conversion tracking their own CRM data their own understanding of what a lead actually costs and what it actually becomes. Not because Google's reporting is unreliable exactly, but because any measurement system that depends entirely on the platform being measured is incomplete by definition.
What Ask Jeeves got right
The original vision was not wrong. A system that understands what you need and finds the answer directly — without requiring you to scan a list of links and make judgements about which one to trust — is genuinely useful. Ask Jeeves understood that in 1996. The technology was not ready.
The technology is ready now. The question is not whether this model arrives — it is already arriving — but what it means for the businesses whose products and services are being represented in those conversations. The agent speaks on your behalf. It uses your content. It applies on your behalf. It costs you money on a pricing model you did not design.
Jeeves always worked for someone else. He just did it very well, and you paid for the service without always knowing exactly what you were getting.
The butler is back. Make sure your house is in order before he starts answering the door.