The core of this chapter is a working definition of a marketing stack, built by hand rather than borrowed from a vendor slide. It names the boxes in plain terms: the CRM as the system of record for contacts and deals, the email or marketing automation platform holding sends, opens, clicks and where each person sits in a nurture sequence, the CMS holding the actual content and its structure, the social scheduler holding the calendar and queued copy, the analytics layer holding behaviour nobody typed into a form, and the spreadsheet everyone forgets to draw but that always exists.
It then walks through how these boxes connect — integrations as agreed field-copies, running over an API, a prebuilt connector, or a webhook — and the three decisions that quietly break every sync: direction (one-way versus two-way), field mapping, and who is allowed to be the source of truth for a given field. A single person's journey from anonymous visitor to CRM record is traced across five tools to show how a cookie identifier and an email address stitch a scattered picture into one person.
From there it lays out the three places AI actually plugs into this picture — inside a tool, between tools, and beside the stack — with a rule of thumb: put AI where the data already lives, and only move data when nothing there can do the job. It closes with a five-step exercise for mapping your own stack on one page, and a warning that AI does not expose dirty or duplicated data, it launders it into polished, confidently wrong output — with duplicate records and quietly manual "integrations" as the usual culprits.
Recent moves get a quick rundown: Salesforce's Agentforce pricing (per-conversation, credit-pack, and per-employee models, plus pay-per-resolution support pricing) and its unified Marketing Cloud; HubSpot's Breeze task agents, mobile workflow approvals, and new AI-search visibility tracking; Google Ads' upcoming shift to exact bid targets and its journey-aware bidding beta; current data on AI Overviews' effect on click-through and citation patterns; a referral-share snapshot showing ChatGPT's dominant but shifting lead over Claude, Gemini, Copilot and Perplexity; and new EU AI Act transparency rules alongside FTC enforcement on AI disclosure.
Last time you stopped re-pasting your brand voice into every new chat. You built one saved, grounded assistant, gave it your style guide and your best past work, and it started sounding like your company instead of like every other company. That was a real win. But look at where that assistant actually sits. It sits beside your business. It knows how you talk. It does not know who bought last week, which of your emails got opened, which blog post is quietly bringing in half your signups, or that the person it is writing to changed jobs in March.
That gap is the whole subject of this chapter, and I want to say the point out loud before we start, because everything else follows from it. AI is only ever as good as the data it can reach. Not the model. Not the prompt. The reach. So the skill you need next is not picking an AI feature off a menu. It is knowing where each fact about a customer actually lives, and which of your systems is allowed to be right about it.
Which means before we wire anything, you have to be able to draw it. You cannot plug AI into a system you cannot draw. So today we draw it — your martech stack, meaning the collection of software you use to do marketing, on one page, by hand, with arrows.
Let me name the boxes in plain words first, and say what each one holds that the others do not. That last part matters, because if two tools held the same data you would only need one of them. The reason a stack has more than one box in it is that each box knows something the rest cannot see.
The first box is your customer relationship management system, your CRM, and the phrase to hold onto is the system of record for contacts and deals. It is where a person becomes an account, a stage, an amount, an owner. Pipeline stage, deal size, contract terms, who on your team owns the relationship, what the last sales call was about. Salesforce and HubSpot are the two names most people run into here. Notice what the CRM does not hold well on its own: it usually knows very little about the eleven months of reading and clicking that happened before anybody talked to sales.
The second box is your email or marketing automation platform — the thing that sends and tracks. Marketo Engage, HubSpot's marketing side, Salesforce's account engagement product. This is where the top of the funnel lives. Every send, open, click and unsubscribe. Every form somebody filled out. Lead scores. And the state of your nurture sequences — a nurture sequence being the timed run of emails somebody gets after they raise a hand, and the platform is holding a little marker for where each person currently sits in that run. That marker exists nowhere else. Your CRM does not know that Maria is on email four of six.
The third box is your content management system, your CMS, where the content itself lives. WordPress, Webflow, Contentful. Published pages, layouts, blog categories and tags, and usually the forms that capture people in the first place. The CMS is the only box that knows the actual words on the page and how the site is structured. That will matter enormously when we get to being visible inside AI answers, because the structure of your content is the thing that gets read.
Fourth is the social scheduler — Sprout Social, Hootsuite, Buffer. It holds the calendar, the queued post copy, and channel-level engagement totals. It is the only place that knows what you are publishing on Thursday.
Fifth is the analytics layer, which tells you what happened. Google Analytics 4, or a product analytics tool like Mixpanel or Amplitude. Sessions, page paths, devices, funnels, how long people stayed. This box holds behaviour that nobody typed into a form. Nobody in the CRM knows that this person read your pricing page three times on a phone.
And the sixth box is the one people leave off the drawing. The spreadsheet. There is always a spreadsheet. Somebody on your team keeps the real list of enterprise accounts, or the real UTM naming conventions, or the actual event registrations, in a sheet with three tabs and a colour code only they understand. Draw it. It is part of your stack whether you like it or not.
Now, how do the boxes connect? Here is the whole idea from the ground up, and I am assuming you have never built one of these.
An integration is two tools agreeing to copy fields to each other. That is it. Somebody, once, sat in a settings screen and said: when this changes over here, put it over there. Sometimes it happens over what is called an application programming interface, an API, which is just the door a piece of software leaves open for other software to talk through. Sometimes it happens through a prebuilt connector the vendors ship because everyone asks for it — the packaged HubSpot-to-Salesforce sync, for instance, or the Marketo-to-Salesforce one. Sometimes it happens through a webhook, which is one tool poking another the instant something happens rather than the second tool asking on a schedule. You do not need to build any of that today. You need to know that behind every arrow on your page there is a decision somebody made.
Three decisions, actually, and they are the three things that break.
The first is direction. A sync has a direction, and one-way versus two-way is a real choice, not a technical footnote. One-way means data flows from here to there and never back. A form fill pushing from your website into your CRM is one-way. Campaign tracking codes flowing into your automation platform and staying there is one-way. Two-way means a change in either system travels to the other. Two-way sounds better. Two-way is also how you get into the most trouble, and I will come back to why.
The second is mapping. The fields have to be paired by name and by type. Your one system calls it Job Title, the other calls it Title. One holds state as free text so anybody can type Calif., the other holds state as a fixed dropdown list that only accepts California. A dropdown has to line up with a dropdown, a date with a date. When they do not line up, the sync does not shout at you. It quietly fails on individual records, one at a time, and the two systems drift apart while both look completely fine on screen.
The third decision is the important one: something has to be the source of truth for each field. If two systems can both write to lifecycle stage, and both have some automation rule of their own about lifecycle stage, they will disagree, and they will each be confident. Real two-way connectors handle this by making you pick — either you name a permanent winner for that field, say the CRM always overrides the marketing platform, or you accept a most-recent-change-wins rule. Both are legitimate. Having never chosen is not.
And what holds all of this together, the spine that lets one human being be the same human being in five different tools, is the email address, in lowercase, because that is the one thing they all collect. Let me walk one person across your page so this stops being abstract.
Somebody lands on your blog from a search. At this moment she is nobody. Your analytics tool and your marketing platform each drop a small file in her browser, a cookie, containing a random identifier — a long meaningless string. That string is not her name. It is a coat check ticket. It lets the system say: this is the same browser that was here on Tuesday. She reads two posts. Nothing has her name yet, and the CRM has no idea she exists.
Then she downloads your guide and types her email address into the form on your CMS. This is the moment everything changes. The system takes that coat check ticket, with its two prior visits attached, and binds it to the email address. Suddenly her history is retroactively hers. Her record appears in your marketing platform with two page views already on it, and the nurture sequence starts.
The marketing platform passes her to the CRM, because she scored high enough, and the CRM creates a record and hands back its own internal identifier — a system-generated key. Now the two tools are holding hands using that key rather than the email address, and that detail is why: when she changes jobs and her email address changes, the link survives, because the tools were never really holding onto the email. They were holding onto the key.
She books a demo. That lives in the CRM. She opens three follow-up emails and clicks one. That lives in the marketing platform. She comes back to the pricing page twice. That lives in analytics. Nobody has the whole person. Everybody has a slice.
Hold that in your head, because now we can talk about AI, and now it will actually make sense.
There are exactly three places AI plugs into this picture. Learn these three and you will be able to place any new AI feature anybody sells you for the next several years, which is why this is the reusable pattern of the chapter rather than a tour of products.
The first place is inside a tool. The platform ships its own AI, and it runs on data that never leaves the building. Your CRM offers predictive lead scoring, or summarises a record, or drafts the sales follow-up. HubSpot's Breeze does prospecting and content drafting. Salesforce has predictive scoring and its agent tooling. Marketo has conversational AI that logs the summary of a chat straight onto the person's record. Your social tool suggests copy and calculates the best time to post from your own past engagement. Your analytics tool flags an anomaly you would not have spotted and estimates who is likely to buy. In every one of those cases the AI is sitting on top of the exact data it needs, already joined, already fresh, and you did not export a thing. Best when the lesson is deployment — when the job is to turn something on and use it well rather than to build anything.
The second place is between tools. You take a hand-off that already exists — one of your arrows — and you put an AI step in the middle of it. Something arrives, AI does one job to it, the result continues on to the next system. That is a genuinely powerful shape and there is a whole craft to building it. We will get there later in this act, when we take on the automation glue properly. For now, all you need is the shape: an arrow with a step in the middle.
The third place is beside the stack. That is your grounded assistant from last chapter. It is not connected to anything. It sees only what you paste into it or what you loaded into its knowledge. Which is exactly, precisely why grounding mattered so much — it has no live line to your business, so the only truth it has is the truth you handed it. It is superb at the work where the material is words and judgement rather than live records: the positioning angle, the brand-voice rewrite, the awkward apology email you need three versions of.
So how do you choose? Rule of thumb, and it is short: put the AI where the data already is, and only move data when nothing there can do the job.
Somebody will object, reasonably: but the AI inside my CRM is not as clever as the assistant I pay for separately. Sometimes true. It is also usually the wrong comparison. A slightly less clever model looking at the correct, current, complete record will beat a brilliant one guessing from a paragraph you pasted in at four in the afternoon. Reach beats raw capability more often than anyone expects. Where the rule genuinely does not hold is when the in-tool feature simply cannot do the task — nobody's CRM is going to write your conference keynote — or when the data you need is spread across three boxes that do not talk to each other, in which case the real fix is not a better assistant, it is an arrow you never built.
Which brings us to the thing to actually do this week. One sitting, one page, real paper if you can stand it.
Now the pitfall, and it is the one that will actually bite you, because it does not look like a failure.
Dirty or duplicated data. AI dropped onto a stack with two competing sources of truth does not surface the mess. It launders it. It takes a wrong field and writes beautiful, confident, on-brand copy off it. It personalises a whole campaign to a job title that went stale eighteen months ago. It scores a lead on engagement data that stopped syncing in the spring, and the score looks exactly as authoritative as a real one. Before AI, bad data produced obviously broken output — the blank merge field, the Dear FIRSTNAME. Now bad data produces polished output. That is worse, not better, because nothing trips your eye.
How it happens is worth knowing, because it is rarely dramatic. Duplicate records are the classic. Somebody merges two duplicates in the CRM, but the merge never travels upstream, so the marketing platform still thinks the deleted record is alive and cheerfully recreates it. Now you have two of her again, and the AI reads whichever one it hits first. Or two teams each built automation on the same field, in different systems, and those rules quietly overwrite each other all day long.
Recognise it by asking the AI where a fact came from. Not whether it is right — where it came from. Which system, which field, when was it last updated. If you get no answer, or a confident answer with no system behind it, you are looking at laundered data and you should not ship what it wrote.
And flag the second tell, which is softer and more common: the manual re-entry step everybody forgot was manual. Somebody exports a list on Fridays and uploads it somewhere. Somebody retypes the event registrations. When you ask, you get told the two tools are integrated, and they are, for some fields, and not for the one you care about. On your page, that is a broken arrow wearing an integration's clothes. The tell is the phrase itself. When you hear we integrated it, ask which fields, which direction, and when it last ran.
Look at your page now. One of those boxes has more marketing-relevant data in it than any other, has AI features already sitting inside it, and is where most people's stack is most confidently wrong. That is the CRM, and next time we open it.
Now, what moved this month.
The biggest item is Salesforce, because the pricing shape tells you what the vendors think they are selling. Its AI catalog is now organised around Agentforce, the agent layer that replaced the older Einstein branding, and it is sold three different ways at once: a per-conversation price of two dollars, a credit pack at five hundred dollars for a hundred thousand credits, and an unmetered per-employee add-on starting at a hundred and twenty-five dollars a user a month. That credit pack is the one to actually understand, because it prices actions, not chats — roughly twenty credits for a standard action, so about ten cents, and about thirty credits for a voice action, so about fifteen. And in mid-year Salesforce added something genuinely new: for customer-facing help agents, pay per resolution, with the billing waived entirely if the conversation escalates to a human or the customer says it went badly. Why you care: agent work is being priced per outcome now, which means your finance conversation changes from seats to volume. Smallest next action is to find out which of the three models your own contract is on before you enable anything.
Alongside that, Salesforce's next-generation Marketing Cloud now sits directly on the core CRM platform with its unified data layer, which brings agent-built campaigns and journey decisioning into the same place your records live. That is the inside-a-tool plug-in point, exactly as we drew it.
HubSpot moved the same direction. Breeze has pushed past drafting into task agents that go and do things — including a data agent that creates properties and enriches records from the open web, and prospecting and customer agents. Two smaller things matter more than they sound. Breeze features are now in the mobile workflows, so approvals happen on a phone. And HubSpot has shipped tracking for AI engine optimisation, meaning visibility inside AI answers, which lands it in your reporting rather than in a side tool. Also new: you can duplicate a campaign by copying its simple workflows, across Starter, Pro and Enterprise. Next action: run the data agent against fifty records and check its enrichment by hand before you let it near five thousand.
On the ad side, the Google Ads bidding change we flagged last time is nearly here — from the seventeenth of August, budget-limited campaigns on target cost per acquisition or target return on ad spend will aim at the exact target you set rather than quietly beating it. If your targets were set loose on the assumption of overperformance, your acquisition cost rises on that date. Two related items are worth your attention: journey-aware bidding is in beta, letting target cost per acquisition learn from mid-funnel events like a qualified lead or a booked demo instead of only the front-end form fill, which is a real gift for anyone with a long sales cycle. And smart bidding exploration has widened across Performance Max and standard shopping, reportedly delivering around an eighteen percent lift in the variety of converting query categories and about nineteen percent more conversions. Next action: reset your targets before the seventeenth, and if you have mid-funnel conversion events already firing, ask for the journey-aware beta.
The standing AI-search visibility picture. AI Overviews appear on somewhere between roughly fourteen and eighteen percent of broad queries, and on half to nearly two thirds of question-shaped ones. When one appears, click-through to the traditional results drops from about fifteen percent to about eight — a fall of nearly half. Roughly six in ten searches now end without a click anywhere. And only about one percent of people click a source link inside an AI answer, which sounds like citation is worthless. It is not, and here is the nuance: about three quarters of the URLs cited in AI Overviews are already ranking in the organic top ten, and the top citation slot gets picked around a third of the time against roughly an eighth for the tenth. Classic ranking still feeds the citation. Meanwhile people recommended a brand by ChatGPT are reportedly two and a half times likelier to visit within a week, and over half of that traffic arrives later as a branded search rather than as a referral you can see. So the visit exists; your report just cannot attribute it.
On referral share among the assistants themselves, one large study across nearly seven million sessions found standalone assistant referrals up roughly tenfold over nineteen months, with ChatGPT taking anywhere from about two thirds to over nine tenths of trackable referrals depending on the industry mix. Claude grew sharply in business-to-business referrals, from around one percent to nearly nineteen, passing Perplexity; Gemini became the steadiest second source; Copilot referrals fell off. Treat all of that as a snapshot, not a standing. Next action: pick one visibility tracker — Similarweb's brand visibility index, or the equivalent inside Semrush or Ahrefs — and get a baseline for five queries your actual customers ask.
And one item from outside the stack that changes what you are allowed to ship. From the second of August, the European Union's transparency rules under its AI Act are fully in force: you must disclose when a person is interacting with an AI system, mark synthetic audio and video in a machine-readable way, and label AI-generated text published without meaningful human editorial oversight. Penalties run up to fifteen million euros or three percent of worldwide turnover. In the United States, the Federal Trade Commission now has a dedicated AI enforcement unit and a maximum civil penalty just over fifty-three thousand dollars per violation, and it expects two disclosures where both apply — the paid relationship and the synthetic testimonial, persona or heavily modified media. Next action: read your own AI Act obligations for the surfaces you actually run, starting with any customer-facing chat, and write down who signs off on labelling. Then come back, because next time we open the CRM box.