How to Get Chosen by Instinct: The 2026 Playbook for Brands and Local Businesses

TL;DR

• Instinct is an invite-only personal assistant you text or call. It runs on a persistent cloud computer with your logins cached, and it is trained to operate a phone and a computer the way a person does.
• That design decides your whole optimisation strategy. Instinct advertises no public merchant catalogue and no publisher programme. It meets your business through the same doors a human assistant would use: the open web, your booking flow, your inbox, your messaging channels and your phone line.
• Since 9 September 2026 it has had its own email address, so it can sign up for services, create accounts and chase support on a user's behalf. Your account wall is no longer the barrier you think it is, and some of your new customer records are about to belong to an agent.
• On 16 September 2026 it announced early access to Instinct Concierge, which places calls to businesses. Meta expanded a Muse outbound-calling beta to US businesses the same day.
• The winning position is unglamorous. Be findable, be quotable, be bookable without a fight, be reachable by phone, and confirm everything in writing.

An investor logged 677 messages to Instinct in five days and got fifteen jobs done, from lowering a Comcast bill to coordinating vendors in India over WhatsApp. In the same period, The Atlantic collected the other half of the ledger: an assistant that cancelled a user's flight while being asked to show the cancellation terms, and an unauthorised restaurant reservation that left a Los Angeles user with a 200 dollar fee.

That is roughly where consumer agents are. Good enough that people hand over real money, rough enough that the failures are memorable. What is not in doubt is the direction. Instinct raised 250 million dollars at a 2.5 billion valuation in August, bringing it to 350 million raised, and by mid-September was reported in talks about a further billion at a 10 billion valuation. Its invite-only user base was reported to have passed 100,000.

Those users are not searching for you. They are asking something to handle it.

This guide is about what "it" sees when it arrives at your business, and what you can change this quarter so the answer is yes.

What is Instinct, and why is it a different problem from a chatbot?

Instinct is built by Spear Street Technology, Inc., a San Francisco company led by Noah Shinn, previously a research scientist at Sierra and first author of the Reflexion paper. Access is invite-only, through a waitlist or an existing member, and it is free during private access with no published price.

You reach it by text message, including iMessage and WhatsApp, or by phone call, all in one ongoing conversation rather than a series of sessions. Behind that conversation, per Vellum's breakdown, it runs on a persistent cloud computer and reaches your connected services with stored credentials. Its terms have the user appoint the service as their agent to enter into agreements, commitments or transactions on their behalf, and say those agreements are binding as if the user had made them directly.

Compare that with the assistant architectures brands have been preparing for.

That last row is the real story. With Instinct, AI search optimisation stops being a marketing project and becomes an operations project. The agent does not care how good your landing page copy is if your booking flow stalls on step three.

How does Instinct choose your business over a competitor?

Nobody has published the retrieval stack, so anyone telling you the ranking factors is guessing. What follows is a model, not documentation, and I will say so plainly where it is inference.

A plausible working model is that the assistant makes the same two decisions a capable human assistant makes. Instinct has published neither the retrieval path nor the ranking, so treat what follows as a way to think about the problem rather than a specification.

First, a shortlist. An agent can only work from what is readable about you: your site, your listings, your reviews, your stated prices and your published availability. That means the groundwork is not different from the work that gets you recommended by ChatGPT. Clear entity data, consistent details across the places you appear, answers written as answers, and facts a machine can lift without inference.

Second, feasibility. This is the part that is new. Once you are on the shortlist, the agent has to settle an operational question that a page of search results usually leaves to the customer: can the job actually be completed here?

If a competitor takes an online booking in four fields and you take one through a portal that demands an account, a verification email and a phone confirmation, expect to lose jobs you should have won. Friction used to cost you a share of converting visitors. Now it can cost you the comparison entirely, before any human sees it.

What are the five doors, and is yours open?

Five is my framework, not Instinct's architecture. It is simply where a general-purpose assistant can touch a business today.

Door one: the open web

The shortlist door, and the cheapest to fix.

• Prices, hours, service areas, policies and availability in real text on the page. A PDF, an image or a widget that arrives late is not impossible for an assistant to handle, it is just less reliable, and reliability is what decides a close call.
• One consistent name, address and phone number everywhere you appear.
• Answers to the actual pre-purchase questions, written in plain sentences, because an assistant can quote a sentence and cannot quote a vibe.
• Deposit, cancellation and change policies stated in words and numbers. When the agent's terms let it make binding commitments and the public reporting already includes a 200 dollar surprise, ambiguity in your terms is a risk you are handing to a stranger.

Door two: the booking flow

The feasibility door, and the one with the highest return.

Run this test yourself. Open your booking or enquiry flow in a private window, with no saved details, and count the steps between intent and confirmation. Then ask an honest question about each step: what does this ask for that only a human could supply?

Common friction, in the order I usually find it:

1. An account requirement before booking. This used to be a wall. Since 9 September, Instinct has had its own email address and can create accounts, sign up for services and chase support itself. So the wall has become a tax: extra steps, more failure points, and a customer record whose email address belongs to an agent rather than to your customer. Offer guest booking anyway.
2. A phone confirmation requirement. You have converted a completed booking into a maybe.
3. Date pickers that only work by mouse gesture. A keyboard-operable calendar gives an agent something to aim at. A bare grid of unlabelled divs gives it far less. Give a text input as well, and test the flow rather than assuming either outcome.
4. CAPTCHA on a legitimate booking path. Sometimes necessary, always a cost. Put it on the abuse path rather than the front door, and know that you are accepting a refusal rate you cannot see.
5. Availability that is only true if you call. If the online calendar is not the real calendar, the agent books the wrong slot or gives up.
6. No confirmation in writing. The agent may be able to verify from an on-screen state or an account record, but a written confirmation survives the session, and it is what your customer will look for when something goes wrong.

Door three: the inbox

Assistants read confirmations. Instinct's privacy policy describes collecting the content of messages, emails and other private communications the user makes accessible, along with audio when enabled and precise location when shared. TechCrunch reported early testers finding emails retained and summarised, in one case after access was revoked. Whatever you think of that, the operational implication is clear: for any customer who grants inbox access, your transactional email may be read and summarised by their assistant.

Send facts as text. One reference number. Dates that cannot be misread, so 14 March 2027 rather than 03/14/27. The change and cancellation link in the body, not buried behind a login. A key detail that exists only inside a banner image will sometimes be read and sometimes not, and you do not get to know which.

Door four: messaging

Instinct lives in iMessage and WhatsApp, and early users have had it coordinate vendors over WhatsApp directly. If your business takes enquiries on WhatsApp, treat that thread like a booking channel rather than a chat: confirm in writing, restate the facts, and avoid the "we will get back to you" holding reply that a human tolerates and an agent treats as a dead end.

Door five: the phone

On 16 September 2026, Instinct announced early access to Concierge. Noah Shinn described it as "a white glove service meant to handle high-touch cases, such as making phone calls, high-end service booking, and more". TechCrunch reported the use cases as restaurants that take no online reservations, dentist cancellation lists and cable billing problems, with wider access planned. The same day, Meta expanded a beta for Muse outbound calls to US businesses, prioritising the users who had already asked for it.

Both are limited rollouts today. Both point the same way, and the last unautomated door into your business is the one that suddenly needs a policy.

What should your policy be when an AI agent calls?

Write it now, in a paragraph, and put it where the front desk can read it. Mine would look like this.

Two things make this policy work. The first is that written confirmation is non-negotiable, because a call is fast, easy to mishear and easy to dispute, and your confirmation is the record both sides can point at. The second is that you should never let a phone call be the only way to transact. Every call an agent has to make is a door you left half shut.

Is this safe to encourage?

Honest answer: the trust picture is unresolved, and you should know it before you optimise for it.

When TechCrunch reported on Instinct on 24 August, the terms then available granted a perpetual and irrevocable licence over user materials, including for model training. Instinct revised its terms two days later. The current terms, last updated 26 August 2026, say the company does not claim ownership of user materials, that it may use them to develop and improve the services including training AI models, that users can opt out of training in settings, and that anything placed in the Vault is used only to provide the service and never for training. That is a meaningful improvement, and it is worth reading the current document rather than the coverage of the old one.

The early failures were real all the same. Claire Vo found emails still being summarised after she disconnected access, and stored in plain text. Alex Cohen deleted his account over how easily the system could be phished. Katie Jacobs Stanton said it "sent an email on my behalf without checking with me first", and that this broke her trust. All of that is documented by TechCrunch.

None of it changes what you should do. You are not endorsing the category. You are making sure that when a customer who has already decided to use one arrives at your business, the transaction completes. Refusing agent-mediated customers is a position some businesses will take. Amazon blocked Meta's Muse, but on narrower grounds: no notice, no consent, an agent that does not identify itself, and concerns about credentials. That is not a blanket rejection of every agent acting for a customer. For a marketplace defending its own buying moment, even the narrow version is defensible. For a restaurant, a clinic, a hotel or a supplier, it mostly means turning away revenue that already chose you.

What you should do is keep the accountable human in the loop where it matters: money, medical detail, anything irreversible. That is not anti-agent. It is the same rule you already apply to a PA calling on behalf of a client.

The thirty-day version

If you only do five things, do these.

1. Book yourself. Anonymous window, no saved details, stopwatch running. Count steps, note every point that needs human judgement or a phone call, and fix the top three.
2. Offer guest booking. Guest checkout, guest enquiry, no account before the transaction. Capture the account afterwards if you want one, and decide now how you will handle records created by an agent's email address.
3. Publish the facts as text. Prices, availability, deposit and cancellation terms, service area, hours. On the page, in sentences, not in a PDF.
4. Rewrite the confirmation email. Plain text facts, one reference, unambiguous dates, a change link that works without a login.
5. Brief the phone. One page, five scenarios, written confirmation every time.

Then measure what you can. No Instinct-specific identifier is documented in the sources I checked, its calls arrive as ordinary calls, and it acts inside a customer's own session, so there is no channel report waiting for you. Use your own records instead. What you can watch is the shape of the traffic: bookings completed unusually fast with no browsing, enquiries written in oddly complete sentences, sign-ups from unfamiliar assistant email domains, callers arranging on someone else's behalf, and confirmations acted on within seconds. Log those. They are the closest thing to an early-warning system for a channel that does not show up in your reports under its own name.

What this is really about

Every wave of search optimisation eventually collapses into the same instruction, and the instruction keeps getting more literal. Be findable. Then be readable. Now, be transactable.

An assistant acting for a customer usually arrives carrying real intent and a specific constraint, which makes it a strong lead. Its budget, its authority and its instructions still need checking, the same as any delegate's. What it does not bring is patience, an ability to guess, or an interest in your brand story.

The businesses that win the next two years will not be the ones with the best copy about how easy they are to work with. They will be the ones that actually are.

Sources

1. TechCrunch, Viral AI startup Instinct has raised 350 million dollars at a 2.5 billion valuation, 26 August 2026.
2. TechCrunch, Instinct's powerful AI assistant is raising privacy and security concerns, 24 August 2026.
3. TechCrunch, Viral AI assistant Instinct now has its own email address, 9 September 2026.
4. TechCrunch, Rival AI agents, Instinct and Meta's Muse, both add the ability to make calls, 17 September 2026.
5. Instinct, Terms of service, last updated 26 August 2026.
6. Instinct, Privacy policy, last updated 26 August 2026.
7. Vellum, Official Instinct breakdown, 2026.
8. PYMNTS, Instinct AI assistant targets 10 billion dollar valuation, 16 September 2026.
9. SiliconANGLE, Consumer-focused AI assistant startup Instinct reportedly raising 250 million dollars, 27 August 2026.
10. The Atlantic, It's all fun and games until you give AI your credit card, September 2026.

Checked against the sources above on 22 September 2026. Instinct is in private access and changing quickly, so treat every figure here as dated rather than settled.