The Age of AI Search: How Dealers Stay Visible and Competitive

Two sessions at the same NADA Show, in the same week, gave South African dealers who watched them almost opposite advice.

The one we wrote up earlier in this series argued that generative engine optimisation is the new frontier and that dealers ignoring it will disappear from AI answers. This one, The Age of AI Search: How Dealers Can Stay Visible and Competitive, presented by Greg Gifford, Chief Operating Officer at SearchLab Digital, spends most of its running time arguing that most of what is being sold as GEO is repackaged SEO, and some of it is snake oil. There is a slide with an actual bottle labelled snake oil on it.

We have written this one up as the counterpoint, because a dealer principal currently being quoted for an “AI visibility” retainer deserves to hear both. The short version of our own position is at the end, and the two sessions agree with each other more than either would enjoy hearing.

The argument in one sentence

AI has changed how people search. It has not changed how many people are buying, and it has not changed what gets you found. As one slide puts it:

“The number of customers looking to buy what you’re selling hasn’t changed, the only thing changing is how they’ll find you.”

Search phrases are getting longer and more conversational. That much everybody agrees on. The disagreement is about what follows from it.

The numbers, and a problem with them

The case against panic rests on four claims made early in the session:

• ChatGPT accounts for about 6 per cent of traffic worldwide, with all query types included.
• ChatGPT shares roughly a 95 per cent audience overlap with Google. The same people are using both.
• Most people who start in ChatGPT go to Google afterwards to fact check.
• SEO still accounts for more than half of revenue for most websites.

We went looking for the sourcing on those and could not find any printed on the slides. The 6 per cent appears as a two-slice pie chart with no title, no legend and no source line. The 95 per cent and the revenue figure appear as sentences.

That is worth saying plainly, because the same session later makes “vague claims are a red flag” one of its own tests for judging a vendor. The direction of all four claims matches what we see in client data, and none of them struck us as implausible. But if you are going to quote one in a management meeting, quote it as this presenter’s figure rather than as a measured fact.

One number in the deck is properly sourced, and it is the interesting one. AI agents succeed on multi-step tasks only about 35 per cent of the time. That comes from Salesforce AI Research’s CRMArena-Pro benchmark, published in May 2025, which found leading agents managing roughly 58 per cent on single-turn tasks and around 35 per cent once the task ran over multiple turns. Worth knowing the benchmark is about business software tasks rather than car buying, but the gap between a one-shot answer and a multi-step job is the point, and it holds.

“Throwing away SEO for GEO is absolute insanity”

This is the spine of the talk, and the argument is made by demonstration rather than assertion. Gifford puts up a run of slides, each headed “Old SEO tactic:”, listing things currently being marketed as generative engine optimisation:

• Create authoritative content featuring original data and expert insights.
• Well-structured, entity-rich content.
• Aim for readability, clarity and easy digestion.
• Anticipate user queries and proactively answer them.

Every one of those is a decade-old content recommendation. That is the whole case, and it is a fair hit. If your prospective GEO supplier’s deliverables list reads like an SEO deliverables list with the acronym changed, you are being sold something you may already be paying for.

One correction, in the interest of the same rigour the session demands of vendors. The slide describes GEO as “a term that was made up six months ago”. The term actually comes from an academic paper published in November 2023, by researchers at Princeton and Georgia Tech among others. The fair version of the complaint is not that the term is new, but that its use as a service category with a monthly fee attached is very new indeed.

The blunter slides are aimed at his own industry rather than at dealers: real SEO as opposed to what OEM-approved providers have supplied for years, a major automotive SEO provider that reportedly dismissed its entire SEO team and replaced them with ChatGPT, and the observation that ChatGPT cannot in fact do SEO. Our favourite line, and one worth repeating to anyone deploying AI in a dealership: AI should really stand for Assume Incorrect.

Grounding: why this is not actually a contradiction

The most useful slide in the deck is a definition, and it is the thing that reconciles this session with the one that appears to disagree with it.

“Grounding: when an AI system does not have the answer in its training data, so it uses search engines to find the answer to a query.”

Ask an assistant which dealership in your suburb has the best service department and it does not know. That is not in its training data at any useful level of local detail, and it changes weekly. So it searches, reads what it finds, and answers from that.

Which means the way to be in the AI answer is to be findable and credible in the search results the model reads. As the slide following it says: they will not be able to find you there if you are not doing SEO.

Why this matters for your dealership: “should we do SEO or GEO” is the wrong question, and both sessions actually answer it the same way. For a local business, generative visibility is downstream of search visibility. The work does not change much. What changes is how you measure whether it is working.

Your traffic is going to fall, and mostly that is fine

Two slides here will save a South African dealer principal an unnecessary argument with their agency.

The first: website traffic will trend down as search behaviour changes, and the traffic you lose is not purchase-intent traffic. The person who used to land on your blog post about tyre pressure and leave is now getting that answer in the AI summary. They were never going to buy a bakkie from you that afternoon.

The second is more practical: Google reduced the amount of data available in Search Console in September, so impressions will be down year on year regardless of how your site is performing. If your dashboard shows a cliff, check the date it starts before you conclude anything.

He is similarly unimpressed by the current crop of AI visibility trackers: good for a hundred-thousand-foot overview, but they do not produce much you can act on. The deeper reason given is one that anybody selling AI rank tracking should have to answer. LLMs are probabilistic. Variability in the answers is not a bug to be engineered around, it is how the technology works. A tool that reports your “ChatGPT rank” as a number is reporting one roll of the dice.

The vendor test

The most immediately useful part of the session is a set of questions to put to anyone selling you AI search services. We would hand this list to any dealer group in Gauteng or the Cape currently reviewing a proposal:

• Can they explain the difference between ChatGPT, Claude and Perplexity? If not, they do not know what they are talking about.
• Do they have actual test results, or claims? Real testing produces real data. Vague claims are a red flag.
• Can they tell you specifically how they monitor and adapt? This field changes weekly, and “we stay on top of it” is not an answer.
• Are they experimenting at your expense? You should not be paying for somebody’s failed experiments.
• Are they ignoring the Google downside? Tactics that chase AI visibility while damaging your organic performance are a bad trade, because organic is still where the conversions are.

Behind all of them is a general scepticism about proprietary AI tooling from marketing suppliers: if the best machine learning researchers have not solved a problem, an automotive marketing vendor has probably not either.

That is a little sweeping, and a supplier does not need to out-research a lab to build something useful on top of one. But as a test of whether a claim deserves your money, it is a good instinct.

What he says to actually do

Strip out the polemic and the recommendations are refreshingly ordinary:

• Answer real questions real people ask. Content that answers genuine queries wins in both Google and AI. He points at the “people also ask” data as the place to find them.
• Make content strategic, not content for its own sake. The negative example he shows is a real dealer page so stuffed with model names and town names that no human could read it.
• Make it chunkable. Clear headings and self-contained sections, so a model can lift a passage and cite it. Which, as he cheerfully points out, is what a decent agency was doing anyway.
• Mention building alongside link building. Being talked about, not just linked to.
• Get reviews on more than one platform. Google alone is not enough now.
• Build real offline relationships. The oldest advice in local marketing, and the hardest to fake.
• Use real images. Avoid stock photography and AI-generated images.
• Use AI to work more efficiently, not to replace people.

Notice how much of that list also appeared in the session we covered earlier. Mention building. Answering the questions buyers actually ask. Structured content a model can quote. Reviews. The two presenters would argue about the label all day and then hand you nearly the same job list.

What changes when you bring this to South Africa

Treat the traffic-share numbers as American. The 6 per cent figure is described as worldwide, but assistant adoption, connectivity and search habits are not evenly distributed, and nobody has published a reliable South African split that we would put in front of a client. Rather than argue about a global number, look at your own Google Analytics referral data for ChatGPT, Perplexity and Gemini. Most South African dealer groups we look at are seeing something real but small, and growing.

Reviews beyond Google means something specific here. The American advice is generic. Ours is not: Hellopeter carries real weight in South Africa in a way it has no equivalent for in the United States, and the motoring marketplaces where buyers already are matter too. If you are only collecting Google reviews, you are only feeding one source.

Local intent is even more local. His point that only locally-intended searches matter to dealers holds, but the unit here is the suburb, not the city. Someone in Fourways is not searching Johannesburg. Your Google Business Profile, your suburb-level pages and your service-area content have to be built at that grain.

The OEM provider problem is worse here. His complaint about manufacturer-approved SEO suppliers will be painfully familiar to any South African dealer working inside an OEM-mandated website platform and marketing programme. You often cannot change supplier. You can still ask the questions in the vendor test above, and you can put the answers in writing to your dealer council.

Offline relationships are a genuine local advantage. Sponsorships, local media, community events and dealer association work all generate the mentions that make you visible to a model. In a smaller market, that is achievable at a scale that would be impossible in a large American metro.

On rands: there are no monetary figures anywhere in this deck to convert. The financial decision it should inform is a different one. Before you add an AI visibility line to a monthly retainer, work out what you are currently paying in rands for SEO, and whether the new line item describes work that is already in it.

Where we would push back

“Who cares” about falling traffic is too breezy. He is right that the lost visits are mostly low intent. But for a dealership whose lead form sits on that website, “traffic will fall and it does not matter” needs a second sentence: watch enquiries, calls and directions requests instead, and if those fall too, it very much does matter. Change the measure, do not abandon measurement.

The sourcing cuts both ways. A session whose central test is “vague claims are a red flag” should not carry three unattributed statistics of its own. We agree with the conclusions. We would just like the footnotes.

And the images. The slide telling dealers to avoid AI-generated images sits in a deck where nearly every illustration is an AI-generated scene of a LEGO figure. We are not throwing stones from a glass house: the photograph at the top of this article was generated with AI too, and we think the advice is right for the pages where it counts. Your vehicles, your premises, your staff and your customers should be photographed, not generated. A decorative image on a blog post is a different question from the picture of the car you are trying to sell.

So which session was right?

Both, mostly, and the fight is narrower than it looks.

The other session is right that buyers are increasingly starting in an assistant, that being mentioned matters more than being linked to, and that a dealership invisible to AI answers has a real problem coming. This session is right that the fix for that is the search work most dealers are still doing badly, that the tooling being sold on top of it is immature, and that nobody should be paying twice for the same deliverables under a newer name.

Our position, for what it is worth: do the SEO properly first, because grounding means it is the mechanism either way. Then measure AI visibility as a check on whether it is working, not as a separate service with its own invoice. If a supplier cannot show you how their GEO deliverables differ from their SEO deliverables, that is your answer.

What to do this month

• Pull your last twelve months of Search Console data and mark where Google changed what it reports, before you read anything into the trend.
• Check your analytics for referrals from ChatGPT, Perplexity and Gemini. Now you have your own number instead of somebody else’s.
• Take your current SEO scope of work and any AI visibility proposal, and put them side by side. Highlight every line that appears in both.
• Run the five vendor questions past your supplier and write down what they say.
• Start collecting reviews somewhere other than Google. Hellopeter is the obvious second.
• Pick your three worst pages, the keyword-stuffed ones nobody could read aloud, and rewrite them to answer one real question each.
• Swap the stock photography on your highest-traffic pages for photographs of your actual premises and staff.

The honest summary

This is the most opinionated session we have covered from the show, and the most useful to a dealer principal holding a proposal. It is also, in places, an agency owner making the case for the service his agency sells, which is worth holding in mind while agreeing with him.

What survives the polemic is a simple and correct model. AI assistants answer local questions by searching, so the way to appear in the answer is to deserve to appear in the search results. Most dealerships have not finished that job. Very few of them will be rescued by buying a new acronym before they do.

If you would like a second opinion on an AI visibility or SEO proposal before you sign it, or an honest read on where your dealership actually stands in both, get in touch. We will tell you if you are already paying for it.

This is the fifth piece in our series on NADA and ATD Show 2026. Read it against 6 Marketing Mistakes and How AI Solves Them, which takes the other side, and alongside Disrupt Yourself Before the Market Disrupts You, Leading With Emotional Intelligence in the AI Age and Driving Service Revenue With AI and CDP Advertising.

Source

• Greg Gifford, Chief Operating Officer, SearchLab Digital, “The Age of AI Search: How Dealers Can Stay Visible and Competitive”, NADA Show 2026, Las Vegas, 3 to 6 February 2026. Quotations are transcribed from the slides.
• The 35 per cent agent success rate is sourced in the deck via a shortlink to Kung-Hsiang Huang and others, “CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions”, Salesforce AI Research, arXiv, May 2025, which reports roughly 58 per cent single-turn and about 35 per cent multi-turn success. The 6 per cent traffic share, the 95 per cent audience overlap and the “more than 50 per cent of revenue” claim carry no source on the slides and are reported here as the presenter’s figures.
• The term GEO was introduced in Pranjal Aggarwal and others, “GEO: Generative Engine Optimization”, arXiv, November 2023, which is why we have corrected the “made up six months ago” line. The dealership shown on stage as an example of poor content is not named here.
• This deck contains no monetary figures, so there was nothing to convert into rands. The South African commentary on review platforms, suburb-level local intent, OEM-mandated suppliers, retainer scope and the four points under “Where we would push back” are ours, not the presenter’s.