Three conversations, three different vocabularies, one question underneath.
The paid search client wants to know why their Google Ads conversions no longer reconcile with their CRM. The content lead wants to know whether to keep publishing at all. The founder doing their own social at nine in the evening wants to know whether any of it is worth the hour.
They think they have three problems. They have one.
Artificial intelligence did not make marketing stop working. It made verification expensive at exactly the moment it made production cheap. Everything that used to be scarce — drafting hours, campaign construction, competent structure — is now nearly free. What became scarce is the ability to say, with a straight face, that a given spend produced a given outcome.
This is the companion piece to our long-form look at the same problem in AI search. Here we take it across paid, content and social.
Paid search: the sentence you can no longer say
Budget approval in most South African businesses runs on one sentence. We spent R50,000, we got forty leads, eight closed.
That sentence is quietly becoming unsayable, and not because anybody did anything wrong.
Start with what Google’s own documentation says about conversions:
“Modeled conversions use data that doesn’t identify individual users to estimate conversions that Google is unable to observe directly.”
“Modeled conversions are reported with the same granularity as observed conversions… In the ‘Conversions’ column, Google reports both modeled and observed conversions.”
Read that twice. Modelled and observed conversions are commingled in a single column, and Google’s documentation does not describe any way for an advertiser to pull them apart. So the honest answer to “how much of my conversion volume is estimated rather than counted?” is: Google does not tell you.
That is not an accusation. Modelling is a reasonable engineering response to consent regimes and browser restrictions that genuinely did remove observable data. But it changes what a conversion number is, and almost nobody has told the person signing the invoice.
Control moved, and one of the controls moved quietly
September 2026 is also when Google began automatically upgrading campaign-level broad match and automatically created assets to AI Max, with Dynamic Search Ads now sunsetting in February 2027. Most advertisers will experience this as something that happened to them rather than something they chose.
There is one detail in Google’s own FAQ worth knowing before you react to it. Negative keywords survive: “Negative keywords will be respected even with AI Max turned on.” But:
“if your campaign is running brand exclusions and AI Max is turned off, brand exclusions will be disabled.”
Turning the feature off disables a brand-safety control. That is not intuitive, it is not prominent in the interface, and we have not seen it covered in the trade press at all. If you are planning to opt out, check your brand exclusions afterwards.
On performance, treat the vendor’s numbers as vendor numbers. Google’s Help Centre cites 14% more conversions or conversion value at similar CPA/ROAS. Its September announcement cites 7%, against a different baseline, footnoted as Google internal data for non-retail advertisers. Both are plausible. Neither is audited, they are not measured against the same thing, and they should never be quoted side by side as a trend.
What Performance Max will and will not show you
Performance Max is not a black box. It is a box with different windows, and the windows do not line up with how most agencies built their reporting. Google’s API documentation is explicit:
“Performance Max campaigns don’t have standard AdGroup and AdGroupAd entities… querying resources such as ad_group or ad_group_ad won’t return any data.”
Campaign search term views, placement views and asset-group reporting all exist. What does not exist is the lever-level diagnosis that agencies spent fifteen years building competence in. When performance drops, you can no longer say this keyword, this placement, this audience. You can say the system decided differently this month.
An agency that cannot diagnose cannot honestly justify a retainer on diagnosis. That is an uncomfortable thing for us to write down, and it is true.
Google has conceded the point itself
The strongest evidence that click-based attribution no longer carries the weight is not an agency opinion. It is that Google shipped the alternative.
Meridian, Google’s open-source Bayesian marketing mix model, became generally available in January 2025. Google’s own framing says traditional MMMs “have historically been unable to fully measure performance media, like Search ads, and AI-powered campaigns.”
When the company selling you the clicks also hands you a free tool to stop trusting click attribution, the debate is over. The remaining question is what you replace it with.
On rising costs, a warning about the numbers
You will read that cost per lead is up nineteen per cent and conversion rates are down fourteen. Check the date on that. The most widely circulating version of those figures is a WordStream benchmark reported in November 2022, and it surfaces in search results today looking entirely current. We expect to see it in other people’s 2026 articles.
More broadly: Google does not publish cost-per-click indices. Every public CPC benchmark is a third-party aggregation of self-selected accounts, usually the vendor’s own customers, with no disclosed weighting. And we could find no primary South African PPC research at all — every local benchmark we traced ended at an agency’s own landing page.
The honest position is that nobody can tell you the market CPC. They can only tell you yours.
Content: the wave crested, and that is the story
The standard narrative is that the internet drowned in AI-generated content. The measured picture is more interesting and considerably more useful.
Graphite analysed 55,400 URLs drawn from Common Crawl between January 2020 and March 2026, classified using three independent AI detectors with false-positive and false-negative rates below two per cent. What they found:
- AI-generated articles overtook human-written articles in volume around November 2024
- Growth then plateaued near fifty per cent and has stayed flat since
- Their hypothesis for the plateau, in their own words: “practitioners found that primarily AI-generated articles do not perform well in search”
The slop wave did not keep rising. It crested and stalled, because the market ran the experiment at scale and the experiment failed. That is a far more encouraging finding than the doom framing, and it has better evidence behind it.
Their caveat is worth carrying: they did not measure traffic, and they did not measure human-edited AI drafts, which they suspect are more common still.
The policy question, answered from the source
Clients ask whether Google will penalise them for using AI. Google’s spam policy defines scaled content abuse as:
“when many pages are generated for the primary purpose of manipulating search rankings and not helping users”
with the illustrative example of using generative tools “to generate many pages without adding value for users.”
The trigger is scale without value. It is not authorship. That is a clean correction to a great deal of nervous folklore, and it comes straight from the documentation.
The number that actually describes the problem
The Content Marketing Institute and MarketingProfs surveyed 1,229 marketers, fielded mid-2025, with disclosed industry and company-size breakdowns. Among B2B respondents:
- 95% use AI-powered applications
- 87% report improved productivity
- 80% report better efficiency
- 58% report improved quality
- 12% report content quality has decreased
And on where the money is going in 2026: AI tools 45%, events 33%, owned media 32%, human resources last, at 9%.
Read those two blocks together. Organisations are buying production capacity roughly five times faster than they are buying the judgement to govern it. The top obstacle named in that survey — unchanged year on year despite universal AI adoption — is “creating content that prompts action,” at 40%.
Volume was never the constraint. It is now cheap enough to prove that conclusively.
What is still worth making by hand
The sorting question that survives is simple: could a language model have produced this from public text?
If yes, it is commodity, and it is exactly the category that saturated and stopped performing. The “what is X” explainer layer is done.
What holds value is what a model cannot synthesise: proprietary data, named client evidence, an opinion with reputational risk attached to it, original research, and the judgement to know which question is worth answering. That is also, not coincidentally, the content that gets cited — by humans and by machines.
Social: the channel where nobody can prove anything
Social is where the gap between required effort and provable return is widest, and everyone involved knows it.
The measured reality on reach is not encouraging. Quid’s 2026 benchmark study — 150 companies from each of eighteen industries, drawn from a database of over 200,000 — puts median engagement at 2.01% on TikTok, 0.30% on Instagram, down seventeen per cent year on year, 0.21% on YouTube and 0.03% on X. Brands are also posting less: Instagram and Facebook posting frequency are both at six-year lows.
Note what that report cannot tell you: it does not cover LinkedIn. And LinkedIn publishes no organic reach benchmarks of its own. Every “LinkedIn organic reach is down X%” figure we could find originated with a scheduling-tool vendor measuring its own users. Treat that entire genre as unusable.
On attribution, HubSpot’s 2026 survey of over 1,100 social professionals found only 37% say it is easy to connect social activity to business outcomes, 69% report pressure to prove ROI, and B2B marketers struggle with attribution notably more than B2C — 41% against 31%.
That is the finding. The inability to measure social is the state of the art, not the failure of a particular team.
The resourcing question nobody enjoys
Underneath “who should own this” sits a genuinely hard problem: the content that works is personal and founder-voiced, and the founder is the least delegable resource in the business.
Constant Contact’s survey of over 5,000 small business owners across five markets found 47% handle all social media themselves. The same research found AI tool adoption among US small businesses rising from 26% in 2023 to 87% by April 2026 — while 74% expect the time they spend on marketing to increase.
Adoption up. Time not returned. That is not a paradox, it is what happens when production gets cheap: the bottleneck moves to judgement, and judgement does not scale by buying a licence.
What this looks like in South Africa
The local picture sharpens all of it.
The 2026 South African Social Media Landscape Report — sixteenth edition, from Ornico, World Wide Worx and Ask Afrika’s Target Group Index — records 73% of South African organisations using AI in marketing and 88% on LinkedIn, the leading corporate platform. It also records that organisations spending under R10,000 a month on social advertising grew from 53% to 60%.
Budgets contracting. Tool adoption climbing. And the barriers respondents named were organisational rather than technological: time, budget, leadership alignment, skills.
Arthur Goldstuck’s framing is the right one: “Artificial intelligence is no longer the story. It is becoming part of the everyday marketing toolkit.”
Peer-reviewed local work agrees on the substance. A study of 44 South African marketing agencies published in SciELO South Africa found high agreement on AI’s usefulness for content production and repetitive-task automation, with a consistent qualitative finding: AI is treated as a framework rather than the end result, with human oversight held essential for accuracy and contextual relevance.
What we would actually do
Five things, in order. None of them are exciting.
One. Find out how much of your reported conversion volume is modelled. You cannot separate it in the Google Ads interface, but you can compare platform-reported conversions against CRM-recorded outcomes over a long enough window to see the shape of the gap. Knowing the size of the gap is worth more than arguing about it.
Two. Measure at the boundary you control. Your CRM, your server logs, your closed revenue. These are first-party, complete, and yours. Platform-reported numbers are inputs to a decision, not the decision.
Three. Stop defending content budgets with traffic. Traffic is falling for structural reasons your content team did not cause and cannot fix. Defend the budget on qualified enquiries and on what the content lets your sales team do. If it produces neither, the problem is the content, not the metric.
Four. Make less and make it unfakeable. Proprietary data, named evidence, real opinion. If a model could have written it from public sources, it is not doing anything for you that a thousand other pages are not already doing.
Five. Decide what social is for, then resource it honestly. If it is founder-led credibility, the founder has to show up and nobody can do that for them. If it is a company page nobody reads, stop, and put the hour somewhere it can be measured.
What we will not tell you
We will not tell you the market cost per click in South Africa, because nobody has measured it and anyone quoting you a figure is re-labelling American data.
We will not tell you that AI Overviews caused your traffic decline, because the best available evidence cannot separate that from the fact that AI Overviews appear more often on queries that were already low-click. Seer Interactive, whose data everyone cites for this, says so themselves.
We will not promise that a given change will produce a given number.
What we will do is tell you which of your numbers are observed and which are estimated, instrument the parts that can be instrumented, and be straight with you about the parts that cannot.
The work still works. Proving it stopped being free — and pretending otherwise is the only real mistake available here.
Related: What Cloudflare’s 15 September change actually does · What the Search Console generative AI report actually measures