I was on a Zoom call last week with a SaaS founder in Berlin.
He shared his screen and typed into ChatGPT: “Who is the market leader in AI compliance software?”
ChatGPT gave him three names. His company wasn’t one of them. The problem? The top answer doesn’t even offer compliance tools. It sells email marketing.
“That’s just wrong,” he said. “But our prospects are quoting it in sales calls.”
This is the reality of search in June 2026. AI answers first, checks facts later — if at all.
Two days before that call, KPMG was forced to pull a major report titled “Redefining excellence in the age of agentic AI” after researchers found it was full of hallucinations. GPTZero checked the footnotes and found only five out of 45 citations actually pointed to the right source. The rest were made up.
That wasn’t a small blog. That was KPMG.
And it’s getting worse, not better. New data shows OpenAI’s o3 and o4-mini models hallucinate at 33% and 48% respectively. For comparison, the previous generation was sitting closer to 16%. Almost half the time, the newest models invent facts.
Marketers feel it. A recent data-backed study found 47% of marketers run into AI errors every single week — outdated info, fabricated stats, or wrong classifications.
This is why people don’t trust AI answers. And it’s why Google is being sued. A German court just ruled Google is liable for AI Overviews that falsely linked two publishers to scams. Google can’t just blame the model anymore.
So where does that leave your website?
Google has a problem. It wants to show AI answers, but it can’t afford to be wrong. Its solution hasn’t changed in 20 years: human votes. A backlink is still the strongest signal that a real editor, on a real website, vouched for you.
AI can fake a paragraph. It can’t fake a link from TechCrunch, or a mention in a trade journal, or a university resource page that has linked to you for three years.
That’s the shift we are seeing with clients since the April 2026 core update. Google specifically targeted “scaled, low-value AI content”. The sites that survived weren’t the ones publishing 100 AI articles a month. They were the ones with 50 to 100 real editorial backlinks pointing to their core pages.
1. We stop chasing DA. We chase real traffic.
A link from a DR 30 site that gets 20,000 real readers a month beats a DR 80 blog with zero traffic. AI models train on what people actually read and cite. We check ahrefs traffic before we pitch.
2. We earn links that correct the record.
For that Berlin client, we didn’t build links to his homepage. We built a single, data-heavy page: “2026 AI Compliance Software Pricing Benchmark (37 Vendors Compared).” We then earned three links to it — one from a cybersecurity newsletter, one from a SaaS review site, and one from a founder interview. Two weeks later, Perplexity started citing that page instead of the wrong competitor.
3. We use expert quotes, not guest posts.
Google’s site reputation abuse update killed parasite guest posts. What works now is digital PR. We put our clients in front of journalists with real data. When a journalist quotes you and links back, that link carries E-E-A-T that AI can’t ignore. It’s a human expert vouching for a human expert.
4. We build links to author pages, not just money pages.
If you want ChatGPT to trust you, it needs to know who you are. We build links to your “About the author” page from podcasts, conference speaker bios, and industry associations. That tells AI: this person is real.
Backlinks in 2026 are not about gaming PageRank. They are about grounding AI.
When a model is about to hallucinate, it looks for consensus. If five trusted sites link to you as the source for “best CRM for manufacturing,” the model is far less likely to invent a competitor.
This is why our link building campaigns now start with one question: “What do we want AI to stop getting wrong about you?” Then we earn the links that fix it.
AI will keep hallucinating. KPMG proved even the biggest firms can’t stop it. Google proved it can’t control it. The only thing you can control is whether real publishers point to you as the truth.
Get the human votes first. The AI citations will follow.