DSRPTERZ Labs is our public record: AI tools and workflows we ran on our own business first, written up with a verdict. Nothing reaches a client until it has survived us, and we publish the failures.
Every write-up follows the same seven steps.
Each one ends in a verdict.
ContinueImproveStop
ProblemWhat is actually slow or expensive, in one sentence.
What we triedThe specific configuration, not the product name.
How we built itSo it can be rebuilt or argued with.
What happenedNumbers where we have them, labelled where we do not.
Where it falls shortThe limits we hit.
What failedThe attempts that did not work, kept in.
What is nextContinue, improve or stop.
Writing
AI tools, myths and workflows, tested on our own campaigns before any of it reaches a client.
Agentic AI in marketing: what it actually does, and what it still gets wrong
Everyone's word of the year. We ran agents on our own campaigns for six weeks. Here is where they held and where they fell over.
The claim
The pitch for 2026 is that agentic AI takes a goal rather than a prompt and executes the multi-step work itself: planning a campaign, writing the variants, adjusting the budget, reporting back. IAB's 2026 outlook found two thirds of buyers focused on agentic AI for ad buying and campaign execution. Gartner's CMO survey puts AI at 15.3% of marketing budgets, with only 30% of CMOs saying they are ready to scale it. That gap is the whole story.
What we ran
Six weeks, our own paid campaigns, three agent configurations. Goal in, multi-step execution out: audience build, variant generation, budget reallocation against a cost-per-result ceiling, weekly summary.
What held
Budget reallocation and reporting. An agent watching a cost-per-result ceiling hourly beats a human checking twice a day, every time. It never forgets, never gets bored, and never talks itself into one more day of a losing ad. The weekly summary was better than ours because it had no incentive to flatter the work.
What fell over
Judgement at the top of the funnel. Given a goal, the agents optimised toward the easiest version of it. Told to lower cost per lead, one shifted spend to an audience that converted cheaply and never bought. The number improved. The business did not. That is not a bug in the model, it is what happens when you hand a proxy metric to something that cannot tell the proxy from the point.
Where it falls short
Agents are excellent at the part of the job with a defined ceiling and a measurable floor. They are poor at deciding what the ceiling should be. They also fail silently: a bad decision looks identical to a good one in the log.
What we do now
Agents run budget pacing, anomaly flagging and reporting on every account. A person sets the goal, picks the metric and signs off anything that changes targeting. Verdict: improve, not continue. We will widen it when we can prove the agent catches its own bad calls.
Published 2026-10-02 by DSRPTERZ Labs. Figures quoted from third parties are as reported at the time of writing.
AI search is eating your traffic. Here is the checklist we run on every site.
Organic traffic is down 15 to 64 percent across categories as answers replace clicks. GEO and AEO are not new disciplines, they are a dozen things you probably have not done.
The shift
Search no longer ends at a results page. People ask ChatGPT, Perplexity, Gemini and Meta AI, get an answer, and never arrive. Reported declines in organic traffic run from 15% to 64% depending on category. Informational content took the worst of it. Transactional queries held up better.
Why most advice is useless
Most GEO advice amounts to write helpful content, which was also the SEO advice for fifteen years. The practical difference is that an answer engine has to be able to read, quote and attribute you, and a surprising number of sites fail at the reading stage.
The checklist
We run twelve points. The five that fail most often: the site needs JavaScript to render its own headings, so a non-rendering crawler sees empty tags. There is no llms.txt. robots.txt blocks GPTBot and ClaudeBot by default, often without anyone deciding to. Claims carry no date or source, so an engine has nothing to attribute. And the page answers a question nobody phrases that way.
The one that surprised people
We audited our own site and found every heading was empty in the HTML — the text was injected by JavaScript for a visual effect. Google renders JS and would have got there eventually. Most AI crawlers do not. We had been invisible to the exact systems we were telling clients to optimise for.
What we do now
Every build ships with server-rendered headings, llms.txt, an explicit allow-list for assistant crawlers, and dated sourcing on every claim. Verdict: continue. It is the cheapest work in the stack and the gap between sites that do it and sites that do not is widening.
Published 2026-09-28 by DSRPTERZ Labs. Figures quoted from third parties are as reported at the time of writing.
The Arabic AI myth: your model does not speak Gulf
Translated English is the single most expensive shortcut in MENA marketing. Here is what breaks, and the two-pass workflow we use instead.
The myth
That a frontier model with multilingual support can produce Arabic marketing copy for the Gulf. It can produce Arabic. That is not the same claim.
What actually breaks
Global models are trained overwhelmingly on English and Modern Standard Arabic. MSA is fine for a government circular and wrong for an Instagram caption. Regional analysis through 2026 has repeatedly identified dialect handling as the most common reason Arabic voice and conversational deployments fail to scale. The copy is grammatical, formal, and reads like a press release in a feed full of people talking normally.
Why it costs real money
Arabic-first content in Saudi Arabia outperforms English on engagement by a wide margin in every benchmark we have seen, with reported lifts from 15-25% on social up to 40-60% on ad engagement, and TikTok conversion gaps reported above 60%. Those numbers are claimed by vendors with an interest in them, so treat the exact figures with suspicion. The direction is consistent enough across independent sources to act on.
The two-pass workflow
Pass one, the model drafts in MSA and we treat it as a brief, not copy. Pass two, a native Gulf speaker rewrites for dialect, register and cultural context. Not a review. A rewrite. The model gets the structure to 80% in minutes; the human does the 20% that makes it sound like a person from Riyadh rather than a person from a dataset.
What we stopped doing
Shipping any Arabic that a native speaker has not rewritten, including captions. We tried a model-only pipeline for low-stakes social and killed it after four weeks. The posts were not wrong. They were invisible.
Verdict
Stop using AI as an Arabic translator. Continue using it as an Arabic drafting assistant with a human finisher. The distinction is the entire job.
Published 2026-09-21 by DSRPTERZ Labs. Figures quoted from third parties are as reported at the time of writing.
Performance Max and Advantage+ are not ad platforms. They are creative briefs.
You no longer choose placements or bids. You feed the machine. Which means the only lever left is the quality of what you feed it.
What changed
Google's Performance Max and Meta's Advantage+ have matured into a different paradigm. You supply creative assets, a product feed, a landing page and an objective. The system handles placement, bidding and targeting. The dials a media buyer used to turn are gone.
The uncomfortable implication
If the algorithm decides placement and bid, then the only thing separating a good account from a bad one is input quality. Your creative, your feed hygiene, your landing page structure. Media buying has quietly become a creative discipline with a spreadsheet attached.
What we test
Asset count and asset variety, measured separately — more assets is not the same as more different assets. Feed completeness, which is boring and matters more than anything else. Landing page to ad message match. And negative signals, because these systems will happily find you cheap conversions from people who will never buy twice.
What we found
The accounts that improved were the ones where we rebuilt the feed, not the ones where we rewrote the ads. That was not what we expected and it is not what the ad copy courses sell.
Verdict
Continue. The practical shift for any team: move budget and hours from bid management into asset production and feed quality. If your media buyer is not also briefing creative, the structure is wrong.
Published 2026-09-14 by DSRPTERZ Labs. Figures quoted from third parties are as reported at the time of writing.
Anti-AI marketing: the counter-trend, and whether it is real
2026 has produced a visible backlash. Brands are advertising that they do not use AI. We looked at whether that is positioning or theatre.
The counter-trend
Alongside the AI adoption curve there is a growing move in the other direction. Some marketers have called 2026 the year of anti-AI marketing, pointing at brands making human-made a claim rather than an assumption. It is an obvious reaction to a feed full of identical generated imagery.
Where it works
In categories where craft is the product. Food, furniture, fashion, anything handmade. Here human-made is not a position against technology, it is a statement about the thing you are selling, and it was true before AI existed.
Where it is theatre
In categories where nobody was ever asking. A logistics company announcing it does not use AI is answering a question its customers have not posed. Worse, it is a claim that ages badly and is almost impossible to verify.
The honest version
The useful position is not anti-AI. It is who decided. Customers do not object to a machine drafting an email. They object to nobody being accountable for what the email says. We put this in writing: AI does research, production, testing and reporting. People read the problem and make every call.
Verdict
Improve. Anti-AI as a slogan is weak. Named human accountability is strong, provable, and survives the trend cycle.
Published 2026-09-07 by DSRPTERZ Labs. Figures quoted from third parties are as reported at the time of writing.
How we test an AI tool before it touches a client: our seven-step method
Most agency AI adoption is a demo, a subscription, and a hope. This is the process we run instead, and the three tools it has killed.
Why a method at all
The AI tooling market rewards demos. A demo is a best case executed by the person who built it. The gap between that and a Tuesday afternoon with real client data is where agency budgets go to die.
The seven steps
Problem — state what is actually slow or expensive, in one sentence, before looking at any tool. What we tried — the specific configuration, not the product name. How we built it — so it can be rebuilt or argued with. What happened — numbers where we have them, labelled where we do not. Where it falls short — the limits we hit. What failed — the attempts that did not work, kept in. What is next — continue, improve or stop.
The rule that makes it work
Everything gets tested on our own business first. Not a sandbox, not a test account. Our campaigns, our reporting, our money. If we are not willing to risk it on ourselves it does not go near a client.
What it has killed
Three tools in the last two quarters. An AI scoring step that disagreed with human judgement more often than it agreed. A meeting-summary pipeline that was accurate and useless, because nobody read the summaries either. And a model-only Arabic pipeline, covered in its own write-up.
What people get wrong
Teams evaluate AI tools on output quality. Output quality is the easy part now. The thing that decides whether a tool survives is whether it fits a workflow someone actually follows when they are busy.
Verdict
Continue. It is slower than buying the subscription and it is the only reason we can tell a client which tools we rejected and why.
Published 2026-08-31 by DSRPTERZ Labs. Figures quoted from third parties are as reported at the time of writing.