The topographic precision of the regenerating retinotectal projection of the goldfish was studied between 18 and 524 days (at 20 degrees C) after optic nerve cut, using retrograde transport of wheatgerm agglutinin conjugated to horseradish peroxidase (WGA-HRP) from one of two standardized tectal injection sites. All labelled ganglion cells in each flat-mounted retina were plotted individually, and their degree of dispersion was assessed by a statistical method based on distance to nearest neighbour. Labelled cells in normal fish were clustered tightly, covering on average only 1.3% of the retina. Early in regeneration (18-28 days) they were widely dispersed, covering up to 75.2%, and they did not begin to form recognizable clusters at appropriate sites until about 35 days after nerve cut. Between 18 and 70 days, the proportion of retina covered by labelled cells fell dramatically, halving about every 14 days. Between 70 and 524 days, no further reduction could be demonstrated: overall, clusters remained significantly larger than normal, though a few individual retinae were virtually normal. Several others, labelled from similar single injections between 56 and 524 days after nerve cut, showed pairs of cell clusters; a sign that persistent errors in topography are common. The very wide initial scatter of labelled cells reflects a striking lack of 'goal-directedness' in regenerative axon growth. Extensive branching in the optic nerve, tract and tectum, for which there is already evidence, must contribute to this. Though uptake of some WGA-HRP by nonsynaptic growth cones cannot be ruled out, other evidence for mislocated functional synapses at early stages encourages us to favour 'trial and error' synapse formation as the likely basis of map refinement.